Skip to content

M
models

PaddlePaddle / models
大约 2 年 前同步成功

通知 232
Star 6828
Fork 2962
M
models
未验证 提交 7430578b 编写于 作者: Y Yibing Liu 提交者： GitHub
浏览文件
操作

Remove models not in release (#1813)

上级 5ddb433e
展开全部 隐藏空白更改
内联 并排
Showing with 0 addition and 137673 deletion
README.md
浏览文件 @ 7430578b
......@@ -59,14 +59,6 @@ PaddlePaddle 提供了丰富的计算单元，使得用户可以采用模块化
[DeepCTR](./fluid/PaddleRec/ctr/README.cn.md)|点击率预估模型|只实现了DeepFM论文中介绍的模型的DNN部分，DeepFM会在其他例子中给出|[DeepFM: A Factorization-Machine based Neural Network for CTR Prediction](https://arxiv.org/abs/1703.04247)
[Multiview-Simnet](./fluid/PaddleRec/multiview_simnet)|个性化推荐模型|基于多元视图，将用户和项目的多个功能视图合并为一个统一模型|[A Multi-View Deep Learning Approach for Cross Domain User Modeling in Recommendation Systems](http://www.microsoft.com/en-us/research/wp-content/uploads/2016/02/frp1159-songA.pdf)
## Other Models
模型|简介|模型优势|参考论文
--|:--:|:--:|:--:
[DeepASR](./fluid/DeepASR/README_cn.md)|语音识别系统|利用Fluid框架完成语音识别中声学模型的配置和训练，并集成 Kaldi 的解码器|-
[DQN](./fluid/DeepQNetwork/README_cn.md)|深度Q网络|value based强化学习算法，第一个成功地将深度学习和强化学习结合起来的模型|[Human-level control through deep reinforcement learning](https://www.nature.com/articles/nature14236)
[DoubleDQN](./fluid/DeepQNetwork/README_cn.md)|DQN的变体|将Double Q的想法应用在DQN上，解决过优化问题|[Font Size: Deep Reinforcement Learning with Double Q-Learning](https://www.aaai.org/ocs/index.php/AAAI/AAAI16/paper/viewPaper/12389)
[DuelingDQN](./fluid/DeepQNetwork/README_cn.md)|DQN的变体|改进了DQN模型，提高了模型的性能|[Dueling Network Architectures for Deep Reinforcement Learning](http://proceedings.mlr.press/v48/wangf16.html)
## License
This tutorial is contributed by [PaddlePaddle](https://github.com/PaddlePaddle/Paddle) and licensed under the [Apache-2.0 license](LICENSE).
......
fluid/AutoDL/LRC/README.md 已删除 100644 → 0
浏览文件 @ 5ddb433e
# LRC Local Rademachar Complexity Regularization
Regularization of Deep Neural Networks(DNNs) for the sake of improving their generalization capability is important and chllenging. This directory contains image classification model based on a novel regularizer rooted in Local Rademacher Complexity (LRC). We appreciate the contribution by [DARTS](https://arxiv.org/abs/1806.09055) for our research. The regularization by LRC and DARTS are combined in this model on CIFAR-10 dataset. Code accompanying the paper
> [An Empirical Study on Regularization of Deep Neural Networks by Local Rademacher Complexity](https://arxiv.org/abs/1902.00873)\
> Yingzhen Yang, Xingjian Li, Jun Huan.\
> _arXiv:1902.00873_.
---
# Table of Contents
- [Installation](#installation)
- [Data preparation](#data-preparation)
- [Training](#training)
## Installation
Running sample code in this directory requires PaddelPaddle Fluid v.1.2.0 and later. If the PaddlePaddle on your device is lower than this version, please follow the instructions in [installation document](http://www.paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/install/index_cn.html#paddlepaddle) and make an update.
## Data preparation
When you want to use the cifar-10 dataset for the first time, you can download the dataset as:
sh ./dataset/download.sh
Please make sure your environment has an internet connection.
The dataset will be downloaded to `dataset/cifar/cifar-10-batches-py` in the same directory as the `train.py`. If automatic download fails, you can download cifar-10-python.tar.gz from https://www.cs.toronto.edu/~kriz/cifar.html and decompress it to the location mentioned above.
## Training
After data preparation, one can start the training step by:
python -u train_mixup.py \
--batch_size=80 \
--auxiliary \
--weight_decay=0.0003 \
--learning_rate=0.025 \
--lrc_loss_lambda=0.7 \
--cutout
- Set ```export CUDA_VISIBLE_DEVICES=0``` to specifiy one GPU to train.
- For more help on arguments:
python train_mixup.py --help
**data reader introduction:**
* Data reader is defined in `reader.py`.
* Reshape the images to 32 * 32.
* In training stage, images are padding to 40 * 40 and cropped randomly to the original size.
* In training stage, images are horizontally random flipped.
* Images are standardized to (0, 1).
* In training stage, cutout images randomly.
* Shuffle the order of the input images during training.
**model configuration:**
* Use auxiliary loss and auxiliary\_weight=0.4.
* Use dropout and drop\_path\_prob=0.2.
* Set lrc\_loss\_lambda=0.7.
**training strategy:**
* Use momentum optimizer with momentum=0.9.
* Weight decay is 0.0003.
* Use cosine decay with init\_lr=0.025.
* Total epoch is 600.
* Use Xaiver initalizer to weight in conv2d, Constant initalizer to weight in batch norm and Normal initalizer to weight in fc.
* Initalize bias in batch norm and fc to zero constant and do not add bias to conv2d.
## Reference
- DARTS: Differentiable Architecture Search [`paper`](https://arxiv.org/abs/1806.09055)
- Differentiable architecture search in PyTorch [`code`](https://github.com/quark0/darts)
fluid/AutoDL/LRC/README_cn.md 已删除 100644 → 0
浏览文件 @ 5ddb433e
# LRC 局部Rademachar复杂度正则化
为了在深度神经网络中提升泛化能力，正则化的选择十分重要也具有挑战性。本目录包括了一种基于局部rademacher复杂度的新型正则（LRC）的图像分类模型。十分感谢[DARTS](https://arxiv.org/abs/1806.09055)模型对本研究提供的帮助。该模型将LRC正则和DARTS网络相结合，在CIFAR-10数据集中得到了很出色的效果。代码和文章一同发布
> [An Empirical Study on Regularization of Deep Neural Networks by Local Rademacher Complexity](https://arxiv.org/abs/1902.00873)\
> Yingzhen Yang, Xingjian Li, Jun Huan.\
> _arXiv:1902.00873_.
---
# 内容
- [安装](#安装)
- [数据准备](#数据准备)
- [模型训练](#模型训练)
## 安装
在当前目录下运行样例代码需要PadddlePaddle Fluid的v.1.2.0或以上的版本。如果你的运行环境中的PaddlePaddle低于此版本，请根据[安装文档](http://www.paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/install/index_cn.html#paddlepaddle)中的说明来更新PaddlePaddle。
## 数据准备
第一次使用CIFAR-10数据集时，您可以通过如果命令下载：
sh ./dataset/download.sh
请确保您的环境有互联网连接。数据会下载到`train.py`同目录下的`dataset/cifar/cifar-10-batches-py`。如果下载失败，您可以自行从https://www.cs.toronto.edu/~kriz/cifar.html上下载cifar-10-python.tar.gz并解压到上述位置。
## 模型训练
数据准备好后，可以通过如下命令开始训练：
python -u train_mixup.py \
--batch_size=80 \
--auxiliary \
--weight_decay=0.0003 \
--learning_rate=0.025 \
--lrc_loss_lambda=0.7 \
--cutout
- 通过设置 ```export CUDA_VISIBLE_DEVICES=0```指定单张GPU训练。
- 可选参数见：
python train_mixup.py --help
**数据读取器说明：**
* 数据读取器定义在`reader.py`中
* 输入图像尺寸统一变换为32 * 32
* 训练时将图像填充为40 * 40然后随机剪裁为原输入图像大小
* 训练时图像随机水平翻转
* 对图像每个像素做归一化处理
* 训练时对图像做随机遮挡
* 训练时对输入图像做随机洗牌
**模型配置：**
* 使用辅助损失，辅助损失权重为0.4
* 使用dropout，随机丢弃率为0.2
* 设置lrc\_loss\_lambda为0.7
**训练策略：**
* 采用momentum优化算法训练，momentum=0.9
* 权重衰减系数为0.0001
* 采用正弦学习率衰减，初始学习率为0.025
* 总共训练600轮
* 对卷积权重采用Xaiver初始化，对batch norm权重采用固定初始化，对全连接层权重采用高斯初始化
* 对batch norm和全连接层偏差采用固定初始化，不对卷积设置偏差
## 引用
- DARTS: Differentiable Architecture Search [`论文`](https://arxiv.org/abs/1806.09055)
- Differentiable Architecture Search in PyTorch [`代码`](https://github.com/quark0/darts)
fluid/AutoDL/LRC/dataset/download.sh 已删除 100644 → 0
浏览文件 @ 5ddb433e
DIR="$( cd "$(dirname "$0")" ; pwd -P )"
cd "$DIR"
mkdir cifar
cd cifar
# Download the data.
echo "Downloading..."
wget https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz
# Extract the data.
echo "Extracting..."
tar zvxf cifar-10-python.tar.gz
fluid/AutoDL/LRC/genotypes.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
# Based on:
# --------------------------------------------------------
# DARTS
# Copyright (c) 2018, Hanxiao Liu.
# Licensed under the Apache License, Version 2.0;
# --------------------------------------------------------
from collections import namedtuple
Genotype = namedtuple('Genotype', 'normal normal_concat reduce reduce_concat')
PRIMITIVES = [
'none', 'max_pool_3x3', 'avg_pool_3x3', 'skip_connect', 'sep_conv_3x3',
'sep_conv_5x5', 'dil_conv_3x3', 'dil_conv_5x5'
]
NASNet = Genotype(
normal=[
('sep_conv_5x5', 1),
('sep_conv_3x3', 0),
('sep_conv_5x5', 0),
('sep_conv_3x3', 0),
('avg_pool_3x3', 1),
('skip_connect', 0),
('avg_pool_3x3', 0),
('avg_pool_3x3', 0),
('sep_conv_3x3', 1),
('skip_connect', 1),
],
normal_concat=[2, 3, 4, 5, 6],
reduce=[
('sep_conv_5x5', 1),
('sep_conv_7x7', 0),
('max_pool_3x3', 1),
('sep_conv_7x7', 0),
('avg_pool_3x3', 1),
('sep_conv_5x5', 0),
('skip_connect', 3),
('avg_pool_3x3', 2),
('sep_conv_3x3', 2),
('max_pool_3x3', 1),
],
reduce_concat=[4, 5, 6], )
AmoebaNet = Genotype(
normal=[
('avg_pool_3x3', 0),
('max_pool_3x3', 1),
('sep_conv_3x3', 0),
('sep_conv_5x5', 2),
('sep_conv_3x3', 0),
('avg_pool_3x3', 3),
('sep_conv_3x3', 1),
('skip_connect', 1),
('skip_connect', 0),
('avg_pool_3x3', 1),
],
normal_concat=[4, 5, 6],
reduce=[
('avg_pool_3x3', 0),
('sep_conv_3x3', 1),
('max_pool_3x3', 0),
('sep_conv_7x7', 2),
('sep_conv_7x7', 0),
('avg_pool_3x3', 1),
('max_pool_3x3', 0),
('max_pool_3x3', 1),
('conv_7x1_1x7', 0),
('sep_conv_3x3', 5),
],
reduce_concat=[3, 4, 6])
DARTS_V1 = Genotype(
normal=[('sep_conv_3x3', 1), ('sep_conv_3x3', 0), ('skip_connect', 0),
('sep_conv_3x3', 1), ('skip_connect', 0), ('sep_conv_3x3', 1),
('sep_conv_3x3', 0), ('skip_connect', 2)],
normal_concat=[2, 3, 4, 5],
reduce=[('max_pool_3x3', 0), ('max_pool_3x3', 1), ('skip_connect', 2),
('max_pool_3x3', 0), ('max_pool_3x3', 0), ('skip_connect', 2),
('skip_connect', 2), ('avg_pool_3x3', 0)],
reduce_concat=[2, 3, 4, 5])
DARTS_V2 = Genotype(
normal=[('sep_conv_3x3', 0), ('sep_conv_3x3', 1), ('sep_conv_3x3', 0),
('sep_conv_3x3', 1), ('sep_conv_3x3', 1), ('skip_connect', 0),
('skip_connect', 0), ('dil_conv_3x3', 2)],
normal_concat=[2, 3, 4, 5],
reduce=[('max_pool_3x3', 0), ('max_pool_3x3', 1), ('skip_connect', 2),
('max_pool_3x3', 1), ('max_pool_3x3', 0), ('skip_connect', 2),
('skip_connect', 2), ('max_pool_3x3', 1)],
reduce_concat=[2, 3, 4, 5])
MY_DARTS = Genotype(
normal=[('sep_conv_3x3', 0), ('skip_connect', 1), ('skip_connect', 0),
('dil_conv_5x5', 1), ('skip_connect', 0), ('sep_conv_3x3', 1),
('skip_connect', 0), ('sep_conv_3x3', 1)],
normal_concat=range(2, 6),
reduce=[('max_pool_3x3', 0), ('max_pool_3x3', 1), ('max_pool_3x3', 0),
('skip_connect', 2), ('max_pool_3x3', 0), ('skip_connect', 2),
('skip_connect', 2), ('skip_connect', 3)],
reduce_concat=range(2, 6))
DARTS = MY_DARTS
fluid/AutoDL/LRC/learning_rate.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
# Based on:
# --------------------------------------------------------
# DARTS
# Copyright (c) 2018, Hanxiao Liu.
# Licensed under the Apache License, Version 2.0;
# --------------------------------------------------------
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import paddle
import paddle.fluid as fluid
import paddle.fluid.layers.ops as ops
from paddle.fluid.layers.learning_rate_scheduler import _decay_step_counter
import math
from paddle.fluid.initializer import init_on_cpu
def cosine_decay(learning_rate, num_epoch, steps_one_epoch):
"""Applies cosine decay to the learning rate.
lr = 0.5 * (math.cos(epoch * (math.pi / 120)) + 1)
"""
global_step = _decay_step_counter()
with init_on_cpu():
decayed_lr = learning_rate * \
(ops.cos((global_step / steps_one_epoch) \
* math.pi / num_epoch) + 1)/2
return decayed_lr
fluid/AutoDL/LRC/model.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
#Licensed under the Apache License, Version 2.0 (the "License");
#you may not use this file except in compliance with the License.
#You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
#Unless required by applicable law or agreed to in writing, software
#distributed under the License is distributed on an "AS IS" BASIS,
#WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
#See the License for the specific language governing permissions and
#limitations under the License.
#
# Based on:
# --------------------------------------------------------
# DARTS
# Copyright (c) 2018, Hanxiao Liu.
# Licensed under the Apache License, Version 2.0;
# --------------------------------------------------------
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import os
import sys
import numpy as np
import time
import functools
import paddle
import paddle.fluid as fluid
from operations import *
class Cell():
def __init__(self, genotype, C_prev_prev, C_prev, C, reduction,
reduction_prev):
print(C_prev_prev, C_prev, C)
if reduction_prev:
self.preprocess0 = functools.partial(FactorizedReduce, C_out=C)
else:
self.preprocess0 = functools.partial(
ReLUConvBN, C_out=C, kernel_size=1, stride=1, padding=0)
self.preprocess1 = functools.partial(
ReLUConvBN, C_out=C, kernel_size=1, stride=1, padding=0)
if reduction:
op_names, indices = zip(*genotype.reduce)
concat = genotype.reduce_concat
else:
op_names, indices = zip(*genotype.normal)
concat = genotype.normal_concat
print(op_names, indices, concat, reduction)
self._compile(C, op_names, indices, concat, reduction)
def _compile(self, C, op_names, indices, concat, reduction):
assert len(op_names) == len(indices)
self._steps = len(op_names) // 2
self._concat = concat
self.multiplier = len(concat)
self._ops = []
for name, index in zip(op_names, indices):
stride = 2 if reduction and index < 2 else 1
op = functools.partial(OPS[name], C=C, stride=stride, affine=True)
self._ops += [op]
self._indices = indices
def forward(self, s0, s1, drop_prob, is_train, name):
self.training = is_train
preprocess0_name = name + 'preprocess0.'
preprocess1_name = name + 'preprocess1.'
s0 = self.preprocess0(s0, name=preprocess0_name)
s1 = self.preprocess1(s1, name=preprocess1_name)
out = [s0, s1]
for i in range(self._steps):
h1 = out[self._indices[2 * i]]
h2 = out[self._indices[2 * i + 1]]
op1 = self._ops[2 * i]
op2 = self._ops[2 * i + 1]
h3 = op1(h1, name=name + '_ops.' + str(2 * i) + '.')
h4 = op2(h2, name=name + '_ops.' + str(2 * i + 1) + '.')
if self.training and drop_prob > 0.:
if h3 != h1:
h3 = fluid.layers.dropout(
h3,
drop_prob,
dropout_implementation='upscale_in_train')
if h4 != h2:
h4 = fluid.layers.dropout(
h4,
drop_prob,
dropout_implementation='upscale_in_train')
s = h3 + h4
out += [s]
return fluid.layers.concat([out[i] for i in self._concat], axis=1)
def AuxiliaryHeadCIFAR(input, num_classes, aux_name='auxiliary_head'):
relu_a = fluid.layers.relu(input)
pool_a = fluid.layers.pool2d(relu_a, 5, 'avg', 3)
conv2d_a = fluid.layers.conv2d(
pool_a,
128,
1,
name=aux_name + '.features.2',
param_attr=ParamAttr(
initializer=Xavier(
uniform=False, fan_in=0),
name=aux_name + '.features.2.weight'),
bias_attr=False)
bn_a_name = aux_name + '.features.3'
bn_a = fluid.layers.batch_norm(
conv2d_a,
act='relu',
name=bn_a_name,
param_attr=ParamAttr(
initializer=Constant(1.), name=bn_a_name + '.weight'),
bias_attr=ParamAttr(
initializer=Constant(0.), name=bn_a_name + '.bias'),
moving_mean_name=bn_a_name + '.running_mean',
moving_variance_name=bn_a_name + '.running_var')
conv2d_b = fluid.layers.conv2d(
bn_a,
768,
2,
name=aux_name + '.features.5',
param_attr=ParamAttr(
initializer=Xavier(
uniform=False, fan_in=0),
name=aux_name + '.features.5.weight'),
bias_attr=False)
bn_b_name = aux_name + '.features.6'
bn_b = fluid.layers.batch_norm(
conv2d_b,
act='relu',
name=bn_b_name,
param_attr=ParamAttr(
initializer=Constant(1.), name=bn_b_name + '.weight'),
bias_attr=ParamAttr(
initializer=Constant(0.), name=bn_b_name + '.bias'),
moving_mean_name=bn_b_name + '.running_mean',
moving_variance_name=bn_b_name + '.running_var')
fc_name = aux_name + '.classifier'
fc = fluid.layers.fc(bn_b,
num_classes,
name=fc_name,
param_attr=ParamAttr(
initializer=Normal(scale=1e-3),
name=fc_name + '.weight'),
bias_attr=ParamAttr(
initializer=Constant(0.), name=fc_name + '.bias'))
return fc
def StemConv(input, C_out, kernel_size, padding):
conv_a = fluid.layers.conv2d(
input,
C_out,
kernel_size,
padding=padding,
param_attr=ParamAttr(
initializer=Xavier(
uniform=False, fan_in=0), name='stem.0.weight'),
bias_attr=False)
bn_a = fluid.layers.batch_norm(
conv_a,
param_attr=ParamAttr(
initializer=Constant(1.), name='stem.1.weight'),
bias_attr=ParamAttr(
initializer=Constant(0.), name='stem.1.bias'),
moving_mean_name='stem.1.running_mean',
moving_variance_name='stem.1.running_var')
return bn_a
class NetworkCIFAR(object):
def __init__(self, C, class_num, layers, auxiliary, genotype):
self.class_num = class_num
self._layers = layers
self._auxiliary = auxiliary
stem_multiplier = 3
self.drop_path_prob = 0
C_curr = stem_multiplier * C
C_prev_prev, C_prev, C_curr = C_curr, C_curr, C
self.cells = []
reduction_prev = False
for i in range(layers):
if i in [layers // 3, 2 * layers // 3]:
C_curr *= 2
reduction = True
else:
reduction = False
cell = Cell(genotype, C_prev_prev, C_prev, C_curr, reduction,
reduction_prev)
reduction_prev = reduction
self.cells += [cell]
C_prev_prev, C_prev = C_prev, cell.multiplier * C_curr
if i == 2 * layers // 3:
C_to_auxiliary = C_prev
def forward(self, init_channel, is_train):
self.training = is_train
self.logits_aux = None
num_channel = init_channel * 3
s0 = StemConv(self.image, num_channel, kernel_size=3, padding=1)
s1 = s0
for i, cell in enumerate(self.cells):
name = 'cells.' + str(i) + '.'
s0, s1 = s1, cell.forward(s0, s1, self.drop_path_prob, is_train,
name)
if i == int(2 * self._layers // 3):
if self._auxiliary and self.training:
self.logits_aux = AuxiliaryHeadCIFAR(s1, self.class_num)
out = fluid.layers.adaptive_pool2d(s1, (1, 1), "avg")
self.logits = fluid.layers.fc(out,
size=self.class_num,
param_attr=ParamAttr(
initializer=Normal(scale=1e-3),
name='classifier.weight'),
bias_attr=ParamAttr(
initializer=Constant(0.),
name='classifier.bias'))
return self.logits, self.logits_aux
def build_input(self, image_shape, batch_size, is_train):
if is_train:
py_reader = fluid.layers.py_reader(
capacity=64,
shapes=[[-1] + image_shape, [-1, 1], [-1, 1], [-1, 1], [-1, 1],
[-1, 1], [-1, batch_size, self.class_num - 1]],
lod_levels=[0, 0, 0, 0, 0, 0, 0],
dtypes=[
"float32", "int64", "int64", "float32", "int32", "int32",
"float32"
],
use_double_buffer=True,
name='train_reader')
else:
py_reader = fluid.layers.py_reader(
capacity=64,
shapes=[[-1] + image_shape, [-1, 1]],
lod_levels=[0, 0],
dtypes=["float32", "int64"],
use_double_buffer=True,
name='test_reader')
return py_reader
def train_model(self, py_reader, init_channels, aux, aux_w, batch_size,
loss_lambda):
self.image, self.ya, self.yb, self.lam, self.label_reshape,\
self.non_label_reshape, self.rad_var = fluid.layers.read_file(py_reader)
self.logits, self.logits_aux = self.forward(init_channels, True)
self.mixup_loss = self.mixup_loss(aux, aux_w)
self.lrc_loss = self.lrc_loss(batch_size)
return self.mixup_loss + loss_lambda * self.lrc_loss
def test_model(self, py_reader, init_channels):
self.image, self.ya = fluid.layers.read_file(py_reader)
self.logits, _ = self.forward(init_channels, False)
prob = fluid.layers.softmax(self.logits, use_cudnn=False)
loss = fluid.layers.cross_entropy(prob, self.ya)
acc_1 = fluid.layers.accuracy(self.logits, self.ya, k=1)
acc_5 = fluid.layers.accuracy(self.logits, self.ya, k=5)
return loss, acc_1, acc_5
def mixup_loss(self, auxiliary, auxiliary_weight):
prob = fluid.layers.softmax(self.logits, use_cudnn=False)
loss_a = fluid.layers.cross_entropy(prob, self.ya)
loss_b = fluid.layers.cross_entropy(prob, self.yb)
loss_a_mean = fluid.layers.reduce_mean(loss_a)
loss_b_mean = fluid.layers.reduce_mean(loss_b)
loss = self.lam * loss_a_mean + (1 - self.lam) * loss_b_mean
if auxiliary:
prob_aux = fluid.layers.softmax(self.logits_aux, use_cudnn=False)
loss_a_aux = fluid.layers.cross_entropy(prob_aux, self.ya)
loss_b_aux = fluid.layers.cross_entropy(prob_aux, self.yb)
loss_a_aux_mean = fluid.layers.reduce_mean(loss_a_aux)
loss_b_aux_mean = fluid.layers.reduce_mean(loss_b_aux)
loss_aux = self.lam * loss_a_aux_mean + (1 - self.lam
) * loss_b_aux_mean
return loss + auxiliary_weight * loss_aux
def lrc_loss(self, batch_size):
y_diff_reshape = fluid.layers.reshape(self.logits, shape=(-1, 1))
label_reshape = fluid.layers.squeeze(self.label_reshape, axes=[1])
non_label_reshape = fluid.layers.squeeze(
self.non_label_reshape, axes=[1])
label_reshape.stop_gradient = True
non_label_reshape.stop_graident = True
y_diff_label_reshape = fluid.layers.gather(y_diff_reshape,
label_reshape)
y_diff_non_label_reshape = fluid.layers.gather(y_diff_reshape,
non_label_reshape)
y_diff_label = fluid.layers.reshape(
y_diff_label_reshape, shape=(-1, batch_size, 1))
y_diff_non_label = fluid.layers.reshape(
y_diff_non_label_reshape,
shape=(-1, batch_size, self.class_num - 1))
y_diff_ = y_diff_non_label - y_diff_label
y_diff_ = fluid.layers.transpose(y_diff_, perm=[1, 2, 0])
rad_var_trans = fluid.layers.transpose(self.rad_var, perm=[1, 2, 0])
rad_y_diff_trans = rad_var_trans * y_diff_
lrc_loss_sum = fluid.layers.reduce_sum(rad_y_diff_trans, dim=[0, 1])
lrc_loss_ = fluid.layers.abs(lrc_loss_sum) / (batch_size *
(self.class_num - 1))
lrc_loss_mean = fluid.layers.reduce_mean(lrc_loss_)
return lrc_loss_mean
fluid/AutoDL/LRC/operations.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
#Licensed under the Apache License, Version 2.0 (the "License");
#you may not use this file except in compliance with the License.
#You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
#Unless required by applicable law or agreed to in writing, software
#distributed under the License is distributed on an "AS IS" BASIS,
#WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
#See the License for the specific language governing permissions and
#limitations under the License.
#
# Based on:
# --------------------------------------------------------
# DARTS
# Copyright (c) 2018, Hanxiao Liu.
# Licensed under the Apache License, Version 2.0;
# --------------------------------------------------------
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import os
import sys
import numpy as np
import time
import paddle
import paddle.fluid as fluid
from paddle.fluid.param_attr import ParamAttr
from paddle.fluid.initializer import Xavier
from paddle.fluid.initializer import Normal
from paddle.fluid.initializer import Constant
OPS = {
'none' : lambda input, C, stride, name, affine: Zero(input, stride, name),
'avg_pool_3x3' : lambda input, C, stride, name, affine: fluid.layers.pool2d(input, 3, 'avg', pool_stride=stride, pool_padding=1, name=name),
'max_pool_3x3' : lambda input, C, stride, name, affine: fluid.layers.pool2d(input, 3, 'max', pool_stride=stride, pool_padding=1, name=name),
'skip_connect' : lambda input,C, stride, name, affine: Identity(input, name) if stride == 1 else FactorizedReduce(input, C, name=name, affine=affine),
'sep_conv_3x3' : lambda input,C, stride, name, affine: SepConv(input, C, C, 3, stride, 1, name=name, affine=affine),
'sep_conv_5x5' : lambda input,C, stride, name, affine: SepConv(input, C, C, 5, stride, 2, name=name, affine=affine),
'sep_conv_7x7' : lambda input,C, stride, name, affine: SepConv(input, C, C, 7, stride, 3, name=name, affine=affine),
'dil_conv_3x3' : lambda input,C, stride, name, affine: DilConv(input, C, C, 3, stride, 2, 2, name=name, affine=affine),
'dil_conv_5x5' : lambda input,C, stride, name, affine: DilConv(input, C, C, 5, stride, 4, 2, name=name, affine=affine),
'conv_7x1_1x7' : lambda input,C, stride, name, affine: SevenConv(input, C, name=name, affine=affine)
}
def ReLUConvBN(input, C_out, kernel_size, stride, padding, name='',
affine=True):
relu_a = fluid.layers.relu(input)
conv2d_a = fluid.layers.conv2d(
relu_a,
C_out,
kernel_size,
stride,
padding,
param_attr=ParamAttr(
initializer=Xavier(
uniform=False, fan_in=0),
name=name + 'op.1.weight'),
bias_attr=False)
if affine:
reluconvbn_out = fluid.layers.batch_norm(
conv2d_a,
param_attr=ParamAttr(
initializer=Constant(1.), name=name + 'op.2.weight'),
bias_attr=ParamAttr(
initializer=Constant(0.), name=name + 'op.2.bias'),
moving_mean_name=name + 'op.2.running_mean',
moving_variance_name=name + 'op.2.running_var')
else:
reluconvbn_out = fluid.layers.batch_norm(
conv2d_a,
param_attr=ParamAttr(
initializer=Constant(1.),
learning_rate=0.,
name=name + 'op.2.weight'),
bias_attr=ParamAttr(
initializer=Constant(0.),
learning_rate=0.,
name=name + 'op.2.bias'),
moving_mean_name=name + 'op.2.running_mean',
moving_variance_name=name + 'op.2.running_var')
return reluconvbn_out
def DilConv(input,
C_in,
C_out,
kernel_size,
stride,
padding,
dilation,
name='',
affine=True):
relu_a = fluid.layers.relu(input)
conv2d_a = fluid.layers.conv2d(
relu_a,
C_in,
kernel_size,
stride,
padding,
dilation,
groups=C_in,
param_attr=ParamAttr(
initializer=Xavier(
uniform=False, fan_in=0),
name=name + 'op.1.weight'),
bias_attr=False,
use_cudnn=False)
conv2d_b = fluid.layers.conv2d(
conv2d_a,
C_out,
1,
param_attr=ParamAttr(
initializer=Xavier(
uniform=False, fan_in=0),
name=name + 'op.2.weight'),
bias_attr=False)
if affine:
dilconv_out = fluid.layers.batch_norm(
conv2d_b,
param_attr=ParamAttr(
initializer=Constant(1.), name=name + 'op.3.weight'),
bias_attr=ParamAttr(
initializer=Constant(0.), name=name + 'op.3.bias'),
moving_mean_name=name + 'op.3.running_mean',
moving_variance_name=name + 'op.3.running_var')
else:
dilconv_out = fluid.layers.batch_norm(
conv2d_b,
param_attr=ParamAttr(
initializer=Constant(1.),
learning_rate=0.,
name=name + 'op.3.weight'),
bias_attr=ParamAttr(
initializer=Constant(0.),
learning_rate=0.,
name=name + 'op.3.bias'),
moving_mean_name=name + 'op.3.running_mean',
moving_variance_name=name + 'op.3.running_var')
return dilconv_out
def SepConv(input,
C_in,
C_out,
kernel_size,
stride,
padding,
name='',
affine=True):
relu_a = fluid.layers.relu(input)
conv2d_a = fluid.layers.conv2d(
relu_a,
C_in,
kernel_size,
stride,
padding,
groups=C_in,
param_attr=ParamAttr(
initializer=Xavier(
uniform=False, fan_in=0),
name=name + 'op.1.weight'),
bias_attr=False,
use_cudnn=False)
conv2d_b = fluid.layers.conv2d(
conv2d_a,
C_in,
1,
param_attr=ParamAttr(
initializer=Xavier(
uniform=False, fan_in=0),
name=name + 'op.2.weight'),
bias_attr=False)
if affine:
bn_a = fluid.layers.batch_norm(
conv2d_b,
param_attr=ParamAttr(
initializer=Constant(1.), name=name + 'op.3.weight'),
bias_attr=ParamAttr(
initializer=Constant(0.), name=name + 'op.3.bias'),
moving_mean_name=name + 'op.3.running_mean',
moving_variance_name=name + 'op.3.running_var')
else:
bn_a = fluid.layers.batch_norm(
conv2d_b,
param_attr=ParamAttr(
initializer=Constant(1.),
learning_rate=0.,
name=name + 'op.3.weight'),
bias_attr=ParamAttr(
initializer=Constant(0.),
learning_rate=0.,
name=name + 'op.3.bias'),
moving_mean_name=name + 'op.3.running_mean',
moving_variance_name=name + 'op.3.running_var')
relu_b = fluid.layers.relu(bn_a)
conv2d_d = fluid.layers.conv2d(
relu_b,
C_in,
kernel_size,
1,
padding,
groups=C_in,
param_attr=ParamAttr(
initializer=Xavier(
uniform=False, fan_in=0),
name=name + 'op.5.weight'),
bias_attr=False,
use_cudnn=False)
conv2d_e = fluid.layers.conv2d(
conv2d_d,
C_out,
1,
param_attr=ParamAttr(
initializer=Xavier(
uniform=False, fan_in=0),
name=name + 'op.6.weight'),
bias_attr=False)
if affine:
sepconv_out = fluid.layers.batch_norm(
conv2d_e,
param_attr=ParamAttr(
initializer=Constant(1.), name=name + 'op.7.weight'),
bias_attr=ParamAttr(
initializer=Constant(0.), name=name + 'op.7.bias'),
moving_mean_name=name + 'op.7.running_mean',
moving_variance_name=name + 'op.7.running_var')
else:
sepconv_out = fluid.layers.batch_norm(
conv2d_e,
param_attr=ParamAttr(
initializer=Constant(1.),
learning_rate=0.,
name=name + 'op.7.weight'),
bias_attr=ParamAttr(
initializer=Constant(0.),
learning_rate=0.,
name=name + 'op.7.bias'),
moving_mean_name=name + 'op.7.running_mean',
moving_variance_name=name + 'op.7.running_var')
return sepconv_out
def SevenConv(input, C_out, stride, name='', affine=True):
relu_a = fluid.layers.relu(input)
conv2d_a = fluid.layers.conv2d(
relu_a,
C_out, (1, 7), (1, stride), (0, 3),
param_attr=ParamAttr(
initializer=Xavier(
uniform=False, fan_in=0),
name=name + 'op.1.weight'),
bias_attr=False)
conv2d_b = fluid.layers.conv2d(
conv2d_a,
C_out, (7, 1), (stride, 1), (3, 0),
param_attr=ParamAttr(
initializer=Xavier(
uniform=False, fan_in=0),
name=name + 'op.2.weight'),
bias_attr=False)
if affine:
out = fluid.layers.batch_norm(
conv2d_b,
param_attr=ParamAttr(
initializer=Constant(1.), name=name + 'op.3.weight'),
bias_attr=ParamAttr(
initializer=Constant(0.), name=name + 'op.3.bias'),
moving_mean_name=name + 'op.3.running_mean',
moving_variance_name=name + 'op.3.running_var')
else:
out = fluid.layers.batch_norm(
conv2d_b,
param_attr=ParamAttr(
initializer=Constant(1.),
learning_rate=0.,
name=name + 'op.3.weight'),
bias_attr=ParamAttr(
initializer=Constant(0.),
learning_rate=0.,
name=name + 'op.3.bias'),
moving_mean_name=name + 'op.3.running_mean',
moving_variance_name=name + 'op.3.running_var')
def Identity(input, name=''):
return input
def Zero(input, stride, name=''):
ones = np.ones(input.shape[-2:])
ones[::stride, ::stride] = 0
ones = fluid.layers.assign(ones)
return input * ones
def FactorizedReduce(input, C_out, name='', affine=True):
relu_a = fluid.layers.relu(input)
conv2d_a = fluid.layers.conv2d(
relu_a,
C_out // 2,
1,
2,
param_attr=ParamAttr(
initializer=Xavier(
uniform=False, fan_in=0),
name=name + 'conv_1.weight'),
bias_attr=False)
h_end = relu_a.shape[2]
w_end = relu_a.shape[3]
slice_a = fluid.layers.slice(relu_a, [2, 3], [1, 1], [h_end, w_end])
conv2d_b = fluid.layers.conv2d(
slice_a,
C_out // 2,
1,
2,
param_attr=ParamAttr(
initializer=Xavier(
uniform=False, fan_in=0),
name=name + 'conv_2.weight'),
bias_attr=False)
out = fluid.layers.concat([conv2d_a, conv2d_b], axis=1)
if affine:
out = fluid.layers.batch_norm(
out,
param_attr=ParamAttr(
initializer=Constant(1.), name=name + 'bn.weight'),
bias_attr=ParamAttr(
initializer=Constant(0.), name=name + 'bn.bias'),
moving_mean_name=name + 'bn.running_mean',
moving_variance_name=name + 'bn.running_var')
else:
out = fluid.layers.batch_norm(
out,
param_attr=ParamAttr(
initializer=Constant(1.),
learning_rate=0.,
name=name + 'bn.weight'),
bias_attr=ParamAttr(
initializer=Constant(0.),
learning_rate=0.,
name=name + 'bn.bias'),
moving_mean_name=name + 'bn.running_mean',
moving_variance_name=name + 'bn.running_var')
return out
fluid/AutoDL/LRC/reader.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
# Copyright (c) 2019 PaddlePaddle Authors. All Rig hts Reserved
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
# Based on:
# --------------------------------------------------------
# DARTS
# Copyright (c) 2018, Hanxiao Liu.
# Licensed under the Apache License, Version 2.0;
# --------------------------------------------------------
"""
CIFAR-10 dataset.
This module will download dataset from
https://www.cs.toronto.edu/~kriz/cifar.html and parse train/test set into
paddle reader creators.
The CIFAR-10 dataset consists of 60000 32x32 colour images in 10 classes,
with 6000 images per class. There are 50000 training images and 10000 test images.
"""
from PIL import Image
from PIL import ImageOps
import numpy as np
import cPickle
import random
import utils
import paddle.fluid as fluid
import time
import os
import functools
import paddle.reader
__all__ = ['train10', 'test10']
image_size = 32
image_depth = 3
half_length = 8
CIFAR_MEAN = [0.4914, 0.4822, 0.4465]
CIFAR_STD = [0.24703233, 0.24348505, 0.26158768]
def generate_reshape_label(label, batch_size, CIFAR_CLASSES=10):
reshape_label = np.zeros((batch_size, 1), dtype='int32')
reshape_non_label = np.zeros(
(batch_size * (CIFAR_CLASSES - 1), 1), dtype='int32')
num = 0
for i in range(batch_size):
label_i = label[i]
reshape_label[i] = label_i + i * CIFAR_CLASSES
for j in range(CIFAR_CLASSES):
if label_i != j:
reshape_non_label[num] = \
j + i * CIFAR_CLASSES
num += 1
return reshape_label, reshape_non_label
def generate_bernoulli_number(batch_size, CIFAR_CLASSES=10):
rcc_iters = 50
rad_var = np.zeros((rcc_iters, batch_size, CIFAR_CLASSES - 1))
for i in range(rcc_iters):
bernoulli_num = np.random.binomial(size=batch_size, n=1, p=0.5)
bernoulli_map = np.array([])
ones = np.ones((CIFAR_CLASSES - 1, 1))
for batch_id in range(batch_size):
num = bernoulli_num[batch_id]
var_id = 2 * ones * num - 1
bernoulli_map = np.append(bernoulli_map, var_id)
rad_var[i] = bernoulli_map.reshape((batch_size, CIFAR_CLASSES - 1))
return rad_var.astype('float32')
def preprocess(sample, is_training, args):
image_array = sample.reshape(3, image_size, image_size)
rgb_array = np.transpose(image_array, (1, 2, 0))
img = Image.fromarray(rgb_array, 'RGB')
if is_training:
# pad and ramdom crop
img = ImageOps.expand(img, (4, 4, 4, 4), fill=0) # pad to 40 * 40 * 3
left_top = np.random.randint(9, size=2) # rand 0 - 8
img = img.crop((left_top[0], left_top[1], left_top[0] + image_size,
left_top[1] + image_size))
if np.random.randint(2):
img = img.transpose(Image.FLIP_LEFT_RIGHT)
img = np.array(img).astype(np.float32)
# per_image_standardization
img_float = img / 255.0
img = (img_float - CIFAR_MEAN) / CIFAR_STD
if is_training and args.cutout:
center = np.random.randint(image_size, size=2)
offset_width = max(0, center[0] - half_length)
offset_height = max(0, center[1] - half_length)
target_width = min(center[0] + half_length, image_size)
target_height = min(center[1] + half_length, image_size)
for i in range(offset_height, target_height):
for j in range(offset_width, target_width):
img[i][j][:] = 0.0
img = np.transpose(img, (2, 0, 1))
return img
def reader_creator_filepath(filename, sub_name, is_training, args):
files = os.listdir(filename)
names = [each_item for each_item in files if sub_name in each_item]
names.sort()
datasets = []
for name in names:
print("Reading file " + name)
batch = cPickle.load(open(filename + name, 'rb'))
data = batch['data']
labels = batch.get('labels', batch.get('fine_labels', None))
assert labels is not None
dataset = zip(data, labels)
datasets.extend(dataset)
random.shuffle(datasets)
def read_batch(datasets, args):
for sample, label in datasets:
im = preprocess(sample, is_training, args)
yield im, [int(label)]
def reader():
batch_data = []
batch_label = []
for data, label in read_batch(datasets, args):
batch_data.append(data)
batch_label.append(label)
if len(batch_data) == args.batch_size:
batch_data = np.array(batch_data, dtype='float32')
batch_label = np.array(batch_label, dtype='int64')
if is_training:
flatten_label, flatten_non_label = \
generate_reshape_label(batch_label, args.batch_size)
rad_var = generate_bernoulli_number(args.batch_size)
mixed_x, y_a, y_b, lam = utils.mixup_data(
batch_data, batch_label, args.batch_size,
args.mix_alpha)
batch_out = [[mixed_x, y_a, y_b, lam, flatten_label, \
flatten_non_label, rad_var]]
yield batch_out
else:
batch_out = [[batch_data, batch_label]]
yield batch_out
batch_data = []
batch_label = []
return reader
def train10(args):
"""
CIFAR-10 training set creator.
It returns a reader creator, each sample in the reader is image pixels in
[0, 1] and label in [0, 9].
:return: Training reader creator
:rtype: callable
"""
return reader_creator_filepath(args.data, 'data_batch', True, args)
def test10(args):
"""
CIFAR-10 test set creator.
It returns a reader creator, each sample in the reader is image pixels in
[0, 1] and label in [0, 9].
:return: Test reader creator.
:rtype: callable
"""
return reader_creator_filepath(args.data, 'test_batch', False, args)
fluid/AutoDL/LRC/run.sh 已删除 100644 → 0
浏览文件 @ 5ddb433e
CUDA_VISIBLE_DEVICES=0 python -u train_mixup.py \
--batch_size=80 \
--auxiliary \
--weight_decay=0.0003 \
--learning_rate=0.025 \
--lrc_loss_lambda=0.7 \
--cutout
fluid/AutoDL/LRC/train_mixup.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
#Licensed under the Apache License, Version 2.0 (the "License");
#you may not use this file except in compliance with the License.
#You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
#Unless required by applicable law or agreed to in writing, software
#distributed under the License is distributed on an "AS IS" BASIS,
#WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
#See the License for the specific language governing permissions and
#limitations under the License.
#
# Based on:
# --------------------------------------------------------
# DARTS
# Copyright (c) 2018, Hanxiao Liu.
# Licensed under the Apache License, Version 2.0;
# --------------------------------------------------------
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from learning_rate import cosine_decay
import numpy as np
import argparse
from model import NetworkCIFAR as Network
import reader
import sys
import os
import time
import logging
import genotypes
import paddle.fluid as fluid
import shutil
import utils
import cPickle as cp
parser = argparse.ArgumentParser("cifar")
parser.add_argument(
'--data',
type=str,
default='./dataset/cifar/cifar-10-batches-py/',
help='location of the data corpus')
parser.add_argument('--batch_size', type=int, default=96, help='batch size')
parser.add_argument(
'--learning_rate', type=float, default=0.025, help='init learning rate')
parser.add_argument('--momentum', type=float, default=0.9, help='momentum')
parser.add_argument(
'--weight_decay', type=float, default=3e-4, help='weight decay')
parser.add_argument(
'--report_freq', type=float, default=50, help='report frequency')
parser.add_argument(
'--epochs', type=int, default=600, help='num of training epochs')
parser.add_argument(
'--init_channels', type=int, default=36, help='num of init channels')
parser.add_argument(
'--layers', type=int, default=20, help='total number of layers')
parser.add_argument(
'--model_path',
type=str,
default='saved_models',
help='path to save the model')
parser.add_argument(
'--auxiliary',
action='store_true',
default=False,
help='use auxiliary tower')
parser.add_argument(
'--auxiliary_weight',
type=float,
default=0.4,
help='weight for auxiliary loss')
parser.add_argument(
'--cutout', action='store_true', default=False, help='use cutout')
parser.add_argument(
'--cutout_length', type=int, default=16, help='cutout length')
parser.add_argument(
'--drop_path_prob', type=float, default=0.2, help='drop path probability')
parser.add_argument('--save', type=str, default='EXP', help='experiment name')
parser.add_argument(
'--arch', type=str, default='DARTS', help='which architecture to use')
parser.add_argument(
'--grad_clip', type=float, default=5, help='gradient clipping')
parser.add_argument(
'--lr_exp_decay',
action='store_true',
default=False,
help='use exponential_decay learning_rate')
parser.add_argument('--mix_alpha', type=float, default=0.5, help='mixup alpha')
parser.add_argument(
'--lrc_loss_lambda', default=0, type=float, help='lrc_loss_lambda')
parser.add_argument(
'--loss_type',
default=1,
type=float,
help='loss_type 0: cross entropy 1: multi margin loss 2: max margin loss')
args = parser.parse_args()
CIFAR_CLASSES = 10
dataset_train_size = 50000
image_size = 32
def main():
image_shape = [3, image_size, image_size]
devices = os.getenv("CUDA_VISIBLE_DEVICES") or ""
devices_num = len(devices.split(","))
logging.info("args = %s", args)
genotype = eval("genotypes.%s" % args.arch)
model = Network(args.init_channels, CIFAR_CLASSES, args.layers,
args.auxiliary, genotype)
steps_one_epoch = dataset_train_size / (devices_num * args.batch_size)
train(model, args, image_shape, steps_one_epoch)
def build_program(main_prog, startup_prog, args, is_train, model, im_shape,
steps_one_epoch):
out = []
with fluid.program_guard(main_prog, startup_prog):
py_reader = model.build_input(im_shape, args.batch_size, is_train)
if is_train:
with fluid.unique_name.guard():
loss = model.train_model(py_reader, args.init_channels,
args.auxiliary, args.auxiliary_weight,
args.batch_size, args.lrc_loss_lambda)
optimizer = fluid.optimizer.Momentum(
learning_rate=cosine_decay(args.learning_rate, \
args.epochs, steps_one_epoch),
regularization=fluid.regularizer.L2Decay(\
args.weight_decay),
momentum=args.momentum)
optimizer.minimize(loss)
out = [py_reader, loss]
else:
with fluid.unique_name.guard():
loss, acc_1, acc_5 = model.test_model(py_reader,
args.init_channels)
out = [py_reader, loss, acc_1, acc_5]
return out
def train(model, args, im_shape, steps_one_epoch):
train_startup_prog = fluid.Program()
test_startup_prog = fluid.Program()
train_prog = fluid.Program()
test_prog = fluid.Program()
train_py_reader, loss_train = build_program(train_prog, train_startup_prog,
args, True, model, im_shape,
steps_one_epoch)
test_py_reader, loss_test, acc_1, acc_5 = build_program(
test_prog, test_startup_prog, args, False, model, im_shape,
steps_one_epoch)
test_prog = test_prog.clone(for_test=True)
place = fluid.CUDAPlace(0)
exe = fluid.Executor(place)
exe.run(train_startup_prog)
exe.run(test_startup_prog)
exec_strategy = fluid.ExecutionStrategy()
exec_strategy.num_threads = 1
train_exe = fluid.ParallelExecutor(
main_program=train_prog,
use_cuda=True,
loss_name=loss_train.name,
exec_strategy=exec_strategy)
train_reader = reader.train10(args)
test_reader = reader.test10(args)
train_py_reader.decorate_paddle_reader(train_reader)
test_py_reader.decorate_paddle_reader(test_reader)
fluid.clip.set_gradient_clip(fluid.clip.GradientClipByNorm(args.grad_clip))
fluid.memory_optimize(fluid.default_main_program())
def save_model(postfix, main_prog):
model_path = os.path.join(args.model_path, postfix)
if os.path.isdir(model_path):
shutil.rmtree(model_path)
fluid.io.save_persistables(exe, model_path, main_program=main_prog)
def test(epoch_id):
test_fetch_list = [loss_test, acc_1, acc_5]
objs = utils.AvgrageMeter()
top1 = utils.AvgrageMeter()
top5 = utils.AvgrageMeter()
test_py_reader.start()
test_start_time = time.time()
step_id = 0
try:
while True:
prev_test_start_time = test_start_time
test_start_time = time.time()
loss_test_v, acc_1_v, acc_5_v = exe.run(
test_prog, fetch_list=test_fetch_list)
objs.update(np.array(loss_test_v), args.batch_size)
top1.update(np.array(acc_1_v), args.batch_size)
top5.update(np.array(acc_5_v), args.batch_size)
if step_id % args.report_freq == 0:
print("Epoch {}, Step {}, acc_1 {}, acc_5 {}, time {}".
format(epoch_id, step_id,
np.array(acc_1_v),
np.array(acc_5_v), test_start_time -
prev_test_start_time))
step_id += 1
except fluid.core.EOFException:
test_py_reader.reset()
print("Epoch {0}, top1 {1}, top5 {2}".format(epoch_id, top1.avg,
top5.avg))
train_fetch_list = [loss_train]
epoch_start_time = time.time()
for epoch_id in range(args.epochs):
model.drop_path_prob = args.drop_path_prob * epoch_id / args.epochs
train_py_reader.start()
epoch_end_time = time.time()
if epoch_id > 0:
print("Epoch {}, total time {}".format(epoch_id - 1, epoch_end_time
- epoch_start_time))
epoch_start_time = epoch_end_time
epoch_end_time
start_time = time.time()
step_id = 0
try:
while True:
prev_start_time = start_time
start_time = time.time()
loss_v, = train_exe.run(
fetch_list=[v.name for v in train_fetch_list])
print("Epoch {}, Step {}, loss {}, time {}".format(epoch_id, step_id, \
np.array(loss_v).mean(), start_time-prev_start_time))
step_id += 1
sys.stdout.flush()
except fluid.core.EOFException:
train_py_reader.reset()
if epoch_id % 50 == 0 or epoch_id == args.epochs - 1:
save_model(str(epoch_id), train_prog)
test(epoch_id)
if __name__ == '__main__':
main()
fluid/AutoDL/LRC/utils.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
# Based on:
# --------------------------------------------------------
# DARTS
# Copyright (c) 2018, Hanxiao Liu.
# Licensed under the Apache License, Version 2.0;
# --------------------------------------------------------
import os
import sys
import time
import math
import numpy as np
def mixup_data(x, y, batch_size, alpha=1.0):
'''Compute the mixup data. Return mixed inputs, pairs of targets, and lambda'''
if alpha > 0.:
lam = np.random.beta(alpha, alpha)
else:
lam = 1.
index = np.random.permutation(batch_size)
mixed_x = lam * x + (1 - lam) * x[index, :]
y_a, y_b = y, y[index]
return mixed_x.astype('float32'), y_a.astype('int64'),\
y_b.astype('int64'), np.array(lam, dtype='float32')
class AvgrageMeter(object):
def __init__(self):
self.reset()
def reset(self):
self.avg = 0
self.sum = 0
self.cnt = 0
def update(self, val, n=1):
self.sum += val * n
self.cnt += n
self.avg = self.sum / self.cnt
fluid/DeepASR/README.md 已删除 100644 → 0
浏览文件 @ 5ddb433e
The minimum PaddlePaddle version needed for the code sample in this directory is the lastest develop branch. If you are on a version of PaddlePaddle earlier than this, [please update your installation](http://www.paddlepaddle.org/docs/develop/documentation/en/build_and_install/pip_install_en.html).
## Deep Automatic Speech Recognition
### Introduction
TBD
### Installation
#### Kaldi
The decoder depends on [kaldi](https://github.com/kaldi-asr/kaldi), install it by flowing its instructions. Then
```shell
export KALDI_ROOT=<absolute path to kaldi>
```
#### Decoder
```shell
git clone https://github.com/PaddlePaddle/models.git
cd models/fluid/DeepASR/decoder
sh setup.sh
```
### Data reprocessing
TBD
### Training
TBD
### Inference & Decoding
TBD
### Question and Contribution
TBD
fluid/DeepASR/README_cn.md 已删除 100644 → 0
浏览文件 @ 5ddb433e
运行本目录下的程序示例需要使用 PaddlePaddle v0.14及以上版本。如果您的 PaddlePaddle 安装版本低于此要求，请按照[安装文档](http://www.paddlepaddle.org/docs/develop/documentation/zh/build_and_install/pip_install_cn.html)中的说明更新 PaddlePaddle 安装版本。
---
DeepASR (Deep Automatic Speech Recognition) 是一个基于PaddlePaddle FLuid与[Kaldi](http://www.kaldi-asr.org)的语音识别系统。其利用Fluid框架完成语音识别中声学模型的配置和训练，并集成 Kaldi 的解码器。旨在方便已对 Kaldi 的较为熟悉的用户实现中声学模型的快速、大规模训练，并利用kaldi完成复杂的语音数据预处理和最终的解码过程。
### 目录
- [模型概览](#model-overview)
- [安装](#installation)
- [数据预处理](#data-reprocessing)
- [模型训练](#training)
- [训练过程中的时间分析](#perf-profiling)
- [预测和解码](#infer-decoding)
- [评估错误率](#scoring-error-rate)
- [Aishell 实例](#aishell-example)
- [欢迎贡献更多的实例](#how-to-contrib)
### 模型概览
DeepASR的声学模型是一个单卷积层加多层层叠LSTMP 的结构，利用卷积来进行初步的特征提取，并用多层的LSTMP来对时序关系进行建模，所用到的损失函数是交叉熵。[LSTMP](https://arxiv.org/abs/1402.1128)(LSTM with recurrent projection layer)是传统 LSTM 的拓展，在 LSTM 的基础上增加了一个映射层，将隐含层映射到较低的维度并输入下一个时间步，这种结构在大为减小 LSTM 的参数规模和计算复杂度的同时还提升了 LSTM 的性能表现。
<p align="center">
<img src="images/lstmp.png" height=240 width=480 hspace='10'/> <br />
图1 LSTMP 的拓扑结构
</p>
### 安装
#### kaldi的安装与设置
DeepASR解码过程中所用的解码器依赖于[Kaldi的安装](https://github.com/kaldi-asr/kaldi)，如环境中无Kaldi, 请`git clone`其源代码，并按给定的命令安装好kaldi，最后设置环境变量`KALDI_ROOT`：
```shell
export KALDI_ROOT=<kaldi的安装路径>
```
#### 解码器的安装
进入解码器源码所在的目录
```shell
cd models/fluid/DeepASR/decoder
```
运行安装脚本
```shell
sh setup.sh
```
编译过程完成即成功地安转了解码器。
### 数据预处理
参考[Kaldi的数据准备流程](http://kaldi-asr.org/doc/data_prep.html)完成音频数据的特征提取和标签对齐
### 声学模型的训练
可以选择在CPU或GPU模式下进行声学模型的训练，例如在GPU模式下的训练
```shell
CUDA_VISIBLE_DEVICES=0,1,2,3 python -u train.py \
--train_feature_lst train_feature.lst \
--train_label_lst train_label.lst \
--val_feature_lst val_feature.lst \
--val_label_lst val_label.lst \
--mean_var global_mean_var \
--parallel
```
其中`train_feature.lst`和`train_label.lst`分别是训练数据集的特征列表文件和标注列表文件，类似的，`val_feature.lst`和`val_label.lst`对应的则是验证集的列表文件。实际训练过程中要正确指定建模单元大小、学习率等重要参数。关于这些参数的说明，请运行
```shell
python train.py --help
```
获取更多信息。
### 训练过程中的时间分析
利用Fluid提供的性能分析工具profiler，可对训练过程进行性能分析，获取网络中operator级别的执行时间
```shell
CUDA_VISIBLE_DEVICES=0 python -u tools/profile.py \
--train_feature_lst train_feature.lst \
--train_label_lst train_label.lst \
--val_feature_lst val_feature.lst \
--val_label_lst val_label.lst \
--mean_var global_mean_var
```
### 预测和解码
在充分训练好声学模型之后，利用训练过程中保存下来的模型checkpoint，可对输入的音频数据进行解码输出，得到声音到文字的识别结果
```
CUDA_VISIBLE_DEVICES=0,1,2,3 python -u infer_by_ckpt.py \
--batch_size 96 \
--checkpoint deep_asr.pass_1.checkpoint \
--infer_feature_lst test_feature.lst \
--infer_label_lst test_label.lst \
--mean_var global_mean_var \
--parallel
```
### 评估错误率
对语音识别系统的评价常用的指标有词错误率(Word Error Rate, WER)和字错误率(Character Error Rate, CER), 在DeepASR中也实现了相关的度量工具，其运行方式为
```
python score_error_rate.py --error_rate_type cer --ref ref.txt --hyp decoding.txt
```
参数`error_rate_type`表示测量错误率的类型，即 WER 或 CER；`ref.txt` 和 `decoding.txt` 分别表示参考文本和实际解码出的文本，它们有着同样的格式：
```
key1 text1
key2 text2
key3 text3
...
```
### Aishell 实例
本节以[Aishell数据集](http://www.aishelltech.com/kysjcp)为例，展示如何完成从数据预处理到解码输出。Aishell是由北京希尔贝克公司所开放的中文普通话语音数据集，时长178小时，包含了400名来自不同口音区域录制者的语音，原始数据可由[openslr](http://www.openslr.org/33)获取。为简化流程，这里提供了已完成预处理的数据集供下载：
```
cd examples/aishell
sh prepare_data.sh
```
其中包括了声学模型的训练数据以及解码过程中所用到的辅助文件等。下载数据完成后，在开始训练之前可对训练过程进行分析
```
sh profile.sh
```
执行训练
```
sh train.sh
```
默认是用4卡GPU进行训练，在实际过程中可根据可用GPU的数目和显存大小对`batch_size`、学习率等参数进行动态调整。训练过程中典型的损失函数和精度的变化趋势如图2所示
<p align="center">
<img src="images/learning_curve.png" height=480 width=640 hspace='10'/> <br />
图2 在Aishell数据集上训练声学模型的学习曲线
</p>
完成模型训练后，即可执行预测识别测试集语音中的文字：
```
sh infer_by_ckpt.sh
```
其中包括了声学模型的预测和解码器的解码输出两个重要的过程。以下是解码输出的样例：
```
...
BAC009S0764W0239 十一 五 期间 我 国 累计 境外 投资 七千亿 美元
BAC009S0765W0140 在 了解 送 方 的 资产 情况 与 需求 之后
BAC009S0915W0291 这 对 苹果 来说 不 是 件 容易 的 事 儿
BAC009S0769W0159 今年 土地 收入 预计 近 四万亿 元
BAC009S0907W0451 由 浦东 商店 作为 掩护
BAC009S0768W0128 土地 交易 可能 随着 供应 淡季 的 到来 而 降温
...
```
每行对应一个输出，均以音频样本的关键字开头，随后是按词分隔的解码出的中文文本。解码完成后运行脚本评估字错误率(CER)
```
sh score_cer.sh
```
其输出类似于如下所示
```
Error rate[cer] = 0.101971 (10683/104765),
total 7176 sentences in hyp, 0 not presented in ref.
```
利用经过20轮左右训练的声学模型，可以在Aishell的测试集上得到CER约10%的识别结果。
### 欢迎贡献更多的实例
DeepASR目前只开放了Aishell实例，我们欢迎用户在更多的数据集上测试完整的训练流程并贡献到这个项目中。
fluid/DeepASR/data_utils/async_data_reader.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
fluid/DeepASR/data_utils/augmentor/tests/__init__.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import data_utils.augmentor.trans_mean_variance_norm as trans_mean_variance_norm
import data_utils.augmentor.trans_add_delta as trans_add_delta
import data_utils.augmentor.trans_splice as trans_splice
fluid/DeepASR/data_utils/augmentor/tests/data/global_mean_var_search26kHr 已删除 100644 → 0
浏览文件 @ 5ddb433e
16.2845556399 11.6891798673
17.21509949 12.3788567902
18.1143704548 14.9912618017
19.2335963752 18.5419556172
19.9266772451 21.2768220522
19.8245737202 21.2347210705
19.5432940972 20.2784036567
19.4631271754 20.2934452329
19.3929919324 20.457971868
19.2924788362 20.3626439234
18.9207244502 19.9196569759
18.7202605641 19.5920276899
18.4844279398 19.2068349019
18.2670948624 18.8716893824
18.0929628855 18.5439666541
17.8428896026 18.0255891747
17.6646850635 17.473764296
17.4955705896 16.8966859471
17.3706720293 16.4294027467
17.2530867792 16.0514717623
17.1304341172 15.7234699057
17.0038353287 15.4344471514
16.902550309 15.1603287337
16.8375590047 14.9304337826
16.816287853 14.9119310513
16.828838265 15.0930023024
16.8602209498 15.3771992423
16.9101763812 15.6897991789
16.9466065143 15.9364556489
16.9486061956 16.0699417826
16.9041374104 16.0796970272
16.8410093699 16.0111444599
16.7045718836 15.7991985601
16.51128489 15.5208920129
16.3253910608 15.2603181921
16.1297317333 14.9499965958
15.903428372 14.5958280409
15.6131718105 14.2709618
15.1395035533 13.9993939893
14.4298229999 13.3841189151
0.0034970565424 0.246184766149
0.00501284154705 0.238484972472
0.00605942680019 0.269064381708
0.00687266156243 0.319479238011
0.00734065019253 0.371947383205
0.00718807218417 0.384426479694
0.00652195540212 0.384676838281
0.00660416525951 0.395543910317
0.00680202057642 0.400803979681
0.00659144183007 0.393228973031
0.00605294530423 0.385021118038
0.00590452969394 0.361763039625
0.00612315374687 0.346777773373
0.00582354093973 0.335802403976
0.00574556002554 0.320733728218
0.00612254485891 0.310153103033
0.00626733043219 0.299854747445
0.00567398408041 0.293353685493
0.00519236700706 0.287668810947
0.00529581474367 0.281479660772
0.00479019484082 0.27451415777
0.00486381039428 0.266294391154
0.00491126372868 0.258105116126
0.00452105305011 0.252926328298
0.00531483334271 0.250910887373
0.00546572110469 0.253302256977
0.00479544857908 0.258484183394
0.00422106426297 0.264582900173
0.00401824135188 0.268467945623
0.0041705465252 0.269699480291
0.00405239564143 0.270406162975
0.0040059737566 0.270407601782
0.00406426729317 0.267951582656
0.00416613791013 0.264543833042
0.00427847607653 0.26247798891
0.00428050903034 0.259635263243
0.00454842971786 0.255829377617
0.00393747552387 0.253802307025
0.00374143688909 0.251011478787
0.00335475310258 0.236543650856
0.000373194755312 0.0419494800709
0.000230909648678 0.0394102370205
0.000150840015851 0.0414956922398
8.44401840771e-05 0.0460502231327
-6.24759314572e-06 0.0528049937739
-8.82957758148e-05 0.055711244886
1.16795791952e-05 0.0563188428833
-1.68716267856e-05 0.0575232763711
-0.000112625308645 0.057979929947
-0.000122619090002 0.0564126233493
1.73569637319e-05 0.05522573909
6.49872782342e-05 0.0507353361334
4.17746389178e-05 0.0479568131253
5.13884475653e-05 0.0461253238047
1.8860115143e-05 0.0436860476919
-5.64317701105e-05 0.042516381059
-0.000136859948115 0.0413574820205
-7.00847019726e-05 0.0409516370727
-5.39392223336e-05 0.040441504085
-9.24897162815e-05 0.0397800398173
4.7104970622e-05 0.039046286243
6.24805896165e-06 0.0380185986602
-2.35272813418e-05 0.036851063786
5.88344154127e-05 0.0361640489242
-8.39162076993e-05 0.0357639427311
-0.000108702805776 0.0358774639538
3.22013961834e-06 0.0363644530435
9.43501518394e-05 0.0370309934774
0.000134406229423 0.0374972993343
3.84007008533e-05 0.037676222515
3.05989328157e-05 0.0379111939182
9.52201629091e-05 0.0380927209106
0.000102126083729 0.0379925358499
6.98628072264e-05 0.0377276252241
4.55782256339e-05 0.0375165468654
4.76370987786e-05 0.0371482526345
-2.24128832709e-05 0.0366810742947
0.000125621306953 0.036628355271
0.000134568666093 0.0364860461759
0.000159858844464 0.0345583593149
fluid/DeepASR/data_utils/augmentor/tests/test_data_trans.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import sys
import unittest
import numpy as np
import data_utils.augmentor.trans_mean_variance_norm as trans_mean_variance_norm
import data_utils.augmentor.trans_add_delta as trans_add_delta
import data_utils.augmentor.trans_splice as trans_splice
import data_utils.augmentor.trans_delay as trans_delay
class TestTransMeanVarianceNorm(unittest.TestCase):
"""unit test for TransMeanVarianceNorm
"""
def setUp(self):
self._file_path = "./data_utils/augmentor/tests/data/" \
"global_mean_var_search26kHr"
def test(self):
feature = np.zeros((2, 120), dtype="float32")
feature.fill(1)
trans = trans_mean_variance_norm.TransMeanVarianceNorm(self._file_path)
(feature1, label1, name) = trans.perform_trans((feature, None, None))
(mean, var) = trans.get_mean_var()
feature_flat1 = feature1.flatten()
feature_flat = feature.flatten()
one = np.ones((1), dtype="float32")
for idx, val in enumerate(feature_flat1):
cur_idx = idx % 120
self.assertAlmostEqual(val, (one[0] - mean[cur_idx]) * var[cur_idx])
class TestTransAddDelta(unittest.TestCase):
"""unit test TestTransAddDelta
"""
def test_regress(self):
"""test regress
"""
feature = np.zeros((14, 120), dtype="float32")
feature[0:5, 0:40].fill(1)
feature[0 + 5, 0:40].fill(1)
feature[1 + 5, 0:40].fill(2)
feature[2 + 5, 0:40].fill(3)
feature[3 + 5, 0:40].fill(4)
feature[8:14, 0:40].fill(4)
trans = trans_add_delta.TransAddDelta()
feature = feature.reshape((14 * 120))
trans._regress(feature, 5 * 120, feature, 5 * 120 + 40, 40, 4, 120)
trans._regress(feature, 5 * 120 + 40, feature, 5 * 120 + 80, 40, 4, 120)
feature = feature.reshape((14, 120))
tmp_feature = feature[5:5 + 4, :]
self.assertAlmostEqual(1.0, tmp_feature[0][0])
self.assertAlmostEqual(0.24, tmp_feature[0][119])
self.assertAlmostEqual(2.0, tmp_feature[1][0])
self.assertAlmostEqual(0.13, tmp_feature[1][119])
self.assertAlmostEqual(3.0, tmp_feature[2][0])
self.assertAlmostEqual(-0.13, tmp_feature[2][119])
self.assertAlmostEqual(4.0, tmp_feature[3][0])
self.assertAlmostEqual(-0.24, tmp_feature[3][119])
def test_perform(self):
"""test perform
"""
feature = np.zeros((4, 40), dtype="float32")
feature[0, 0:40].fill(1)
feature[1, 0:40].fill(2)
feature[2, 0:40].fill(3)
feature[3, 0:40].fill(4)
trans = trans_add_delta.TransAddDelta()
(feature, label, name) = trans.perform_trans((feature, None, None))
self.assertAlmostEqual(feature.shape[0], 4)
self.assertAlmostEqual(feature.shape[1], 120)
self.assertAlmostEqual(1.0, feature[0][0])
self.assertAlmostEqual(0.24, feature[0][119])
self.assertAlmostEqual(2.0, feature[1][0])
self.assertAlmostEqual(0.13, feature[1][119])
self.assertAlmostEqual(3.0, feature[2][0])
self.assertAlmostEqual(-0.13, feature[2][119])
self.assertAlmostEqual(4.0, feature[3][0])
self.assertAlmostEqual(-0.24, feature[3][119])
class TestTransSplict(unittest.TestCase):
"""unit test Test TransSplict
"""
def test_perfrom(self):
feature = np.zeros((8, 10), dtype="float32")
for i in xrange(feature.shape[0]):
feature[i, :].fill(i)
trans = trans_splice.TransSplice()
(feature, label, name) = trans.perform_trans((feature, None, None))
self.assertEqual(feature.shape[1], 110)
for i in xrange(8):
nzero_num = 5 - i
cur_val = 0.0
if nzero_num < 0:
cur_val = i - 5 - 1
for j in xrange(11):
if j <= nzero_num:
for k in xrange(10):
self.assertAlmostEqual(feature[i][j * 10 + k], cur_val)
else:
if cur_val < 7:
cur_val += 1.0
for k in xrange(10):
self.assertAlmostEqual(feature[i][j * 10 + k], cur_val)
class TestTransDelay(unittest.TestCase):
"""unittest TransDelay
"""
def test_perform(self):
label = np.zeros((10, 1), dtype="int64")
for i in xrange(10):
label[i][0] = i
trans = trans_delay.TransDelay(5)
(_, label, _) = trans.perform_trans((None, label, None))
for i in xrange(5):
self.assertAlmostEqual(label[i + 5][0], i)
for i in xrange(5):
self.assertAlmostEqual(label[i][0], 0)
if __name__ == '__main__':
unittest.main()
fluid/DeepASR/data_utils/augmentor/trans_add_delta.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import numpy as np
import math
import copy
class TransAddDelta(object):
""" add delta of feature data
trans feature for shape(a, b) to shape(a, b * 3)
Attributes:
_norder(int):
_window(int):
"""
def __init__(self, norder=2, nwindow=2):
""" init construction
Args:
norder: default 2
nwindow: default 2
"""
self._norder = norder
self._nwindow = nwindow
def perform_trans(self, sample):
""" add delta for feature
trans feature shape from (a,b) to (a, b * 3)
Args:
sample(object,tuple): contain feature numpy and label numpy
Returns:
(feature, label, name)
"""
(feature, label, name) = sample
frame_dim = feature.shape[1]
d_frame_dim = frame_dim * 3
head_filled = 5
tail_filled = 5
mat = np.zeros(
(feature.shape[0] + head_filled + tail_filled, d_frame_dim),
dtype="float32")
#copy first frame
for i in xrange(head_filled):
np.copyto(mat[i, 0:frame_dim], feature[0, :])
np.copyto(mat[head_filled:head_filled + feature.shape[0], 0:frame_dim],
feature[:, :])
# copy last frame
for i in xrange(head_filled + feature.shape[0], mat.shape[0], 1):
np.copyto(mat[i, 0:frame_dim], feature[feature.shape[0] - 1, :])
nframe = feature.shape[0]
start = head_filled
tmp_shape = mat.shape
mat = mat.reshape((tmp_shape[0] * tmp_shape[1]))
self._regress(mat, start * d_frame_dim, mat,
start * d_frame_dim + frame_dim, frame_dim, nframe,
d_frame_dim)
self._regress(mat, start * d_frame_dim + frame_dim, mat,
start * d_frame_dim + 2 * frame_dim, frame_dim, nframe,
d_frame_dim)
mat.shape = tmp_shape
return (mat[head_filled:mat.shape[0] - tail_filled, :], label, name)
def _regress(self, data_in, start_in, data_out, start_out, size, n, step):
""" regress
Args:
data_in: in data
start_in: start index of data_in
data_out: out data
start_out: start index of data_out
size: frame dimentional
n: frame num
step: 3 * (frame num)
Returns:
None
"""
sigma_t2 = 0.0
delta_window = self._nwindow
for t in xrange(1, delta_window + 1):
sigma_t2 += t * t
sigma_t2 *= 2.0
for i in xrange(n):
fp1 = start_in
fp2 = start_out
for j in xrange(size):
back = fp1
forw = fp1
sum = 0.0
for t in xrange(1, delta_window + 1):
back -= step
forw += step
sum += t * (data_in[forw] - data_in[back])
data_out[fp2] = sum / sigma_t2
fp1 += 1
fp2 += 1
start_in += step
start_out += step
fluid/DeepASR/data_utils/augmentor/trans_delay.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import numpy as np
import math
class TransDelay(object):
""" Delay label, and copy first label value in the front.
Attributes:
_delay_time : the delay frame num of label
"""
def __init__(self, delay_time):
"""init construction
Args:
delay_time : the delay frame num of label
"""
self._delay_time = delay_time
def perform_trans(self, sample):
"""
Args:
sample(object):input sample, contain feature numpy and label numpy, sample name list
Returns:
(feature, label, name)
"""
(feature, label, name) = sample
shape = label.shape
assert len(shape) == 2
label[self._delay_time:shape[0]] = label[0:shape[0] - self._delay_time]
for i in xrange(self._delay_time):
label[i][0] = label[self._delay_time][0]
return (feature, label, name)
fluid/DeepASR/data_utils/augmentor/trans_mean_variance_norm.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import numpy as np
import math
class TransMeanVarianceNorm(object):
""" normalization of mean variance for feature data
Attributes:
_mean(numpy.array): the feature mean vector
_var(numpy.array): the feature variance
"""
def __init__(self, snorm_path):
"""init construction
Args:
snorm_path: the path of mean and variance
"""
self._mean = None
self._var = None
self._load_norm(snorm_path)
def _load_norm(self, snorm_path):
""" load mean var file
Args:
snorm_path(str):the file path
"""
lLines = open(snorm_path).readlines()
nLen = len(lLines)
self._mean = np.zeros((nLen), dtype="float32")
self._var = np.zeros((nLen), dtype="float32")
self._nLen = nLen
for nidx, l in enumerate(lLines):
s = l.split()
assert len(s) == 2
self._mean[nidx] = float(s[0])
self._var[nidx] = 1.0 / math.sqrt(float(s[1]))
if self._var[nidx] > 100000.0:
self._var[nidx] = 100000.0
def get_mean_var(self):
""" get mean and var
Args:
Returns:
(mean, var)
"""
return (self._mean, self._var)
def perform_trans(self, sample):
""" feature = (feature - mean) * var
Args:
sample(object):input sample, contain feature numpy and label numpy
Returns:
(feature, label, name)
"""
(feature, label, name) = sample
shape = feature.shape
assert len(shape) == 2
nfeature_len = shape[0] * shape[1]
assert nfeature_len % self._nLen == 0
ncur_idx = 0
feature = feature.reshape((nfeature_len))
while ncur_idx < nfeature_len:
block = feature[ncur_idx:ncur_idx + self._nLen]
block = (block - self._mean) * self._var
feature[ncur_idx:ncur_idx + self._nLen] = block
ncur_idx += self._nLen
feature = feature.reshape(shape)
return (feature, label, name)
fluid/DeepASR/data_utils/augmentor/trans_splice.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import numpy as np
import math
class TransSplice(object):
""" copy feature context to construct new feature
expand feature data from shape (frame_num, frame_dim)
to shape (frame_num, frame_dim * 11)
Attributes:
_nleft_context(int): copy left context number
_nright_context(int): copy right context number
"""
def __init__(self, nleft_context=5, nright_context=5):
""" init construction
Args:
nleft_context(int):
nright_context(int):
"""
self._nleft_context = nleft_context
self._nright_context = nright_context
def perform_trans(self, sample):
""" copy feature context
Args:
sample(object): input sample(feature, label)
Return:
(feature, label, name)
"""
(feature, label, name) = sample
nframe_num = feature.shape[0]
nframe_dim = feature.shape[1]
nnew_frame_dim = nframe_dim * (
self._nleft_context + self._nright_context + 1)
mat = np.zeros(
(nframe_num + self._nleft_context + self._nright_context,
nframe_dim),
dtype="float32")
ret = np.zeros((nframe_num, nnew_frame_dim), dtype="float32")
#copy left
for i in xrange(self._nleft_context):
mat[i, :] = feature[0, :]
#copy middle
mat[self._nleft_context:self._nleft_context +
nframe_num, :] = feature[:, :]
#copy right
for i in xrange(self._nright_context):
mat[i + self._nleft_context + nframe_num, :] = feature[-1, :]
mat = mat.reshape(mat.shape[0] * mat.shape[1])
ret = ret.reshape(ret.shape[0] * ret.shape[1])
for i in xrange(nframe_num):
np.copyto(ret[i * nnew_frame_dim:(i + 1) * nnew_frame_dim],
mat[i * nframe_dim:i * nframe_dim + nnew_frame_dim])
ret = ret.reshape((nframe_num, nnew_frame_dim))
return (ret, label, name)
fluid/DeepASR/data_utils/util.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import sys
from six import reraise
from tblib import Traceback
import numpy as np
def to_lodtensor(data, place):
"""convert tensor to lodtensor
"""
seq_lens = [len(seq) for seq in data]
cur_len = 0
lod = [cur_len]
for l in seq_lens:
cur_len += l
lod.append(cur_len)
flattened_data = numpy.concatenate(data, axis=0).astype("int64")
flattened_data = flattened_data.reshape([len(flattened_data), 1])
res = fluid.LoDTensor()
res.set(flattened_data, place)
res.set_lod([lod])
return res
def split_infer_result(infer_seq, lod):
infer_batch = []
for i in xrange(0, len(lod[0]) - 1):
infer_batch.append(infer_seq[lod[0][i]:lod[0][i + 1]])
return infer_batch
class CriticalException(Exception):
pass
def suppress_signal(signo, stack_frame):
pass
def suppress_complaints(verbose, notify=None):
def decorator_maker(func):
def suppress_warpper(*args, **kwargs):
try:
func(*args, **kwargs)
except:
et, ev, tb = sys.exc_info()
if notify is not None:
notify(except_type=et, except_value=ev, traceback=tb)
if verbose == 1 or isinstance(ev, CriticalException):
reraise(et, ev, Traceback(tb).as_traceback())
return suppress_warpper
return decorator_maker
class ForceExitWrapper(object):
def __init__(self, exit_flag):
self._exit_flag = exit_flag
@suppress_complaints(verbose=0)
def __call__(self, *args, **kwargs):
self._exit_flag.value = True
def __eq__(self, flag):
return self._exit_flag.value == flag
fluid/DeepASR/decoder/.gitignore 已删除 100644 → 0
浏览文件 @ 5ddb433e
ThreadPool
build
post_latgen_faster_mapped.so
pybind11
fluid/DeepASR/decoder/post_latgen_faster_mapped.cc 已删除 100644 → 0
浏览文件 @ 5ddb433e
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "post_latgen_faster_mapped.h"
#include <limits>
#include "ThreadPool.h"
using namespace kaldi;
typedef kaldi::int32 int32;
using fst::SymbolTable;
using fst::Fst;
using fst::StdArc;
Decoder::Decoder(std::string trans_model_in_filename,
std::string word_syms_filename,
std::string fst_in_filename,
std::string logprior_in_filename,
size_t beam_size,
kaldi::BaseFloat acoustic_scale) {
const char *usage =
"Generate lattices using neural net model.\n"
"Usage: post-latgen-faster-mapped [options] <trans-model> "
"<fst-in|fsts-rspecifier> <logprior> <posts-rspecifier>"
" <lattice-wspecifier> [ <words-wspecifier> [<alignments-wspecifier>] "
"]\n";
ParseOptions po(usage);
allow_partial = false;
this->acoustic_scale = acoustic_scale;
config.Register(&po);
int32 beam = 11;
po.Register("acoustic-scale",
&acoustic_scale,
"Scaling factor for acoustic likelihoods");
po.Register("word-symbol-table",
&word_syms_filename,
"Symbol table for words [for debug output]");
po.Register("allow-partial",
&allow_partial,
"If true, produce output even if end state was not reached.");
int argc = 2;
char *argv[] = {(char *)"post-latgen-faster-mapped",
(char *)("--beam=" + std::to_string(beam_size)).c_str()};
po.Read(argc, argv);
std::ifstream is_logprior(logprior_in_filename);
logprior.Read(is_logprior, false);
{
bool binary;
Input ki(trans_model_in_filename, &binary);
this->trans_model.Read(ki.Stream(), binary);
}
this->determinize = config.determinize_lattice;
this->word_syms = NULL;
if (word_syms_filename != "") {
if (!(word_syms = fst::SymbolTable::ReadText(word_syms_filename))) {
KALDI_ERR << "Could not read symbol table from file "
<< word_syms_filename;
}
}
// Input FST is just one FST, not a table of FSTs.
this->decode_fst = fst::ReadFstKaldiGeneric(fst_in_filename);
kaldi::LatticeFasterDecoder *decoder =
new LatticeFasterDecoder(*decode_fst, config);
decoder_pool.emplace_back(decoder);
std::string lattice_wspecifier =
"ark:|gzip -c > mapped_decoder_data/lat.JOB.gz";
if (!(determinize ? compact_lattice_writer.Open(lattice_wspecifier)
: lattice_writer.Open(lattice_wspecifier)))
KALDI_ERR << "Could not open table for writing lattices: "
<< lattice_wspecifier;
words_writer = new Int32VectorWriter("");
alignment_writer = new Int32VectorWriter("");
}
Decoder::~Decoder() {
if (!this->word_syms) delete this->word_syms;
delete this->decode_fst;
for (size_t i = 0; i < decoder_pool.size(); ++i) {
delete decoder_pool[i];
}
delete words_writer;
delete alignment_writer;
}
void Decoder::decode_from_file(std::string posterior_rspecifier,
size_t num_processes) {
try {
double tot_like = 0.0;
kaldi::int64 frame_count = 0;
// int num_success = 0, num_fail = 0;
KALDI_ASSERT(ClassifyRspecifier(fst_in_filename, NULL, NULL) ==
kNoRspecifier);
SequentialBaseFloatMatrixReader posterior_reader("ark:" +
posterior_rspecifier);
Timer timer;
timer.Reset();
double elapsed = 0.0;
for (size_t n = decoder_pool.size(); n < num_processes; ++n) {
kaldi::LatticeFasterDecoder *decoder =
new LatticeFasterDecoder(*decode_fst, config);
decoder_pool.emplace_back(decoder);
}
elapsed = timer.Elapsed();
ThreadPool thread_pool(num_processes);
while (!posterior_reader.Done()) {
timer.Reset();
std::vector<std::future<std::string>> que;
for (size_t i = 0; i < num_processes && !posterior_reader.Done(); ++i) {
std::string utt = posterior_reader.Key();
Matrix<BaseFloat> &loglikes(posterior_reader.Value());
que.emplace_back(thread_pool.enqueue(std::bind(
&Decoder::decode_internal, this, decoder_pool[i], utt, loglikes)));
posterior_reader.Next();
}
timer.Reset();
for (size_t i = 0; i < que.size(); ++i) {
std::cout << que[i].get() << std::endl;
}
}
} catch (const std::exception &e) {
std::cerr << e.what();
}
}
inline kaldi::Matrix<kaldi::BaseFloat> vector2kaldi_mat(
const std::vector<std::vector<kaldi::BaseFloat>> &log_probs) {
size_t num_frames = log_probs.size();
size_t dim_label = log_probs[0].size();
kaldi::Matrix<kaldi::BaseFloat> loglikes(
num_frames, dim_label, kaldi::kSetZero, kaldi::kStrideEqualNumCols);
for (size_t i = 0; i < num_frames; ++i) {
memcpy(loglikes.Data() + i * dim_label,
log_probs[i].data(),
sizeof(kaldi::BaseFloat) * dim_label);
}
return loglikes;
}
std::vector<std::string> Decoder::decode_batch(
std::vector<std::string> keys,
const std::vector<std::vector<std::vector<kaldi::BaseFloat>>>
&log_probs_batch,
size_t num_processes) {
ThreadPool thread_pool(num_processes);
std::vector<std::string> decoding_results; //(keys.size(), "");
for (size_t n = decoder_pool.size(); n < num_processes; ++n) {
kaldi::LatticeFasterDecoder *decoder =
new LatticeFasterDecoder(*decode_fst, config);
decoder_pool.emplace_back(decoder);
}
size_t index = 0;
while (index < keys.size()) {
std::vector<std::future<std::string>> res_in_que;
for (size_t t = 0; t < num_processes && index < keys.size(); ++t) {
kaldi::Matrix<kaldi::BaseFloat> loglikes =
vector2kaldi_mat(log_probs_batch[index]);
res_in_que.emplace_back(
thread_pool.enqueue(std::bind(&Decoder::decode_internal,
this,
decoder_pool[t],
keys[index],
loglikes)));
index++;
}
for (size_t i = 0; i < res_in_que.size(); ++i) {
decoding_results.emplace_back(res_in_que[i].get());
}
}
return decoding_results;
}
std::string Decoder::decode(
std::string key,
const std::vector<std::vector<kaldi::BaseFloat>> &log_probs) {
kaldi::Matrix<kaldi::BaseFloat> loglikes = vector2kaldi_mat(log_probs);
return decode_internal(decoder_pool[0], key, loglikes);
}
std::string Decoder::decode_internal(
LatticeFasterDecoder *decoder,
std::string key,
kaldi::Matrix<kaldi::BaseFloat> &loglikes) {
if (loglikes.NumRows() == 0) {
KALDI_WARN << "Zero-length utterance: " << key;
// num_fail++;
}
KALDI_ASSERT(loglikes.NumCols() == logprior.Dim());
loglikes.ApplyLog();
loglikes.AddVecToRows(-1.0, logprior);
DecodableMatrixScaledMapped matrix_decodable(
trans_model, loglikes, acoustic_scale);
double like;
return this->DecodeUtteranceLatticeFaster(
decoder, matrix_decodable, key, &like);
}
std::string Decoder::DecodeUtteranceLatticeFaster(
LatticeFasterDecoder *decoder,
DecodableInterface &decodable, // not const but is really an input.
std::string utt,
double *like_ptr) { // puts utterance's like in like_ptr on success.
using fst::VectorFst;
std::string ret = utt + ' ';
if (!decoder->Decode(&decodable)) {
KALDI_WARN << "Failed to decode file " << utt;
return ret;
}
if (!decoder->ReachedFinal()) {
if (allow_partial) {
KALDI_WARN << "Outputting partial output for utterance " << utt
<< " since no final-state reached\n";
} else {
KALDI_WARN << "Not producing output for utterance " << utt
<< " since no final-state reached and "
<< "--allow-partial=false.\n";
return ret;
}
}
double likelihood;
LatticeWeight weight;
int32 num_frames;
{ // First do some stuff with word-level traceback...
VectorFst<LatticeArc> decoded;
if (!decoder->GetBestPath(&decoded))
// Shouldn't really reach this point as already checked success.
KALDI_ERR << "Failed to get traceback for utterance " << utt;
std::vector<int32> alignment;
std::vector<int32> words;
GetLinearSymbolSequence(decoded, &alignment, &words, &weight);
num_frames = alignment.size();
// if (alignment_writer->IsOpen()) alignment_writer->Write(utt, alignment);
if (word_syms != NULL) {
for (size_t i = 0; i < words.size(); i++) {
std::string s = word_syms->Find(words[i]);
ret += s + ' ';
}
}
likelihood = -(weight.Value1() + weight.Value2());
}
// Get lattice, and do determinization if requested.
Lattice lat;
decoder->GetRawLattice(&lat);
if (lat.NumStates() == 0)
KALDI_ERR << "Unexpected problem getting lattice for utterance " << utt;
fst::Connect(&lat);
if (determinize) {
CompactLattice clat;
if (!DeterminizeLatticePhonePrunedWrapper(
trans_model,
&lat,
decoder->GetOptions().lattice_beam,
&clat,
decoder->GetOptions().det_opts))
KALDI_WARN << "Determinization finished earlier than the beam for "
<< "utterance " << utt;
// We'll write the lattice without acoustic scaling.
if (acoustic_scale != 0.0)
fst::ScaleLattice(fst::AcousticLatticeScale(1.0 / acoustic_scale), &clat);
// disable output lattice temporarily
// compact_lattice_writer.Write(utt, clat);
} else {
// We'll write the lattice without acoustic scaling.
if (acoustic_scale != 0.0)
fst::ScaleLattice(fst::AcousticLatticeScale(1.0 / acoustic_scale), &lat);
// lattice_writer.Write(utt, lat);
}
return ret;
}
fluid/DeepASR/decoder/post_latgen_faster_mapped.h 已删除 100644 → 0
浏览文件 @ 5ddb433e
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include <string>
#include <vector>
#include "base/kaldi-common.h"
#include "base/timer.h"
#include "decoder/decodable-matrix.h"
#include "decoder/decoder-wrappers.h"
#include "fstext/kaldi-fst-io.h"
#include "hmm/transition-model.h"
#include "tree/context-dep.h"
#include "util/common-utils.h"
class Decoder {
public:
Decoder(std::string trans_model_in_filename,
std::string word_syms_filename,
std::string fst_in_filename,
std::string logprior_in_filename,
size_t beam_size,
kaldi::BaseFloat acoustic_scale);
~Decoder();
// Interface to accept the scores read from specifier and print
// the decoding results directly
void decode_from_file(std::string posterior_rspecifier,
size_t num_processes = 1);
// Accept the scores of one utterance and return the decoding result
std::string decode(
std::string key,
const std::vector<std::vector<kaldi::BaseFloat>> &log_probs);
// Accept the scores of utterances in batch and return the decoding results
std::vector<std::string> decode_batch(
std::vector<std::string> key,
const std::vector<std::vector<std::vector<kaldi::BaseFloat>>>
&log_probs_batch,
size_t num_processes = 1);
private:
// For decoding one utterance
std::string decode_internal(kaldi::LatticeFasterDecoder *decoder,
std::string key,
kaldi::Matrix<kaldi::BaseFloat> &loglikes);
std::string DecodeUtteranceLatticeFaster(kaldi::LatticeFasterDecoder *decoder,
kaldi::DecodableInterface &decodable,
std::string utt,
double *like_ptr);
fst::SymbolTable *word_syms;
fst::Fst<fst::StdArc> *decode_fst;
std::vector<kaldi::LatticeFasterDecoder *> decoder_pool;
kaldi::Vector<kaldi::BaseFloat> logprior;
kaldi::TransitionModel trans_model;
kaldi::LatticeFasterDecoderConfig config;
kaldi::CompactLatticeWriter compact_lattice_writer;
kaldi::LatticeWriter lattice_writer;
kaldi::Int32VectorWriter *words_writer;
kaldi::Int32VectorWriter *alignment_writer;
bool binary;
bool determinize;
kaldi::BaseFloat acoustic_scale;
bool allow_partial;
};
fluid/DeepASR/decoder/pybind.cc 已删除 100644 → 0
浏览文件 @ 5ddb433e
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include <pybind11/pybind11.h>
#include <pybind11/stl.h>
#include "post_latgen_faster_mapped.h"
namespace py = pybind11;
PYBIND11_MODULE(post_latgen_faster_mapped, m) {
m.doc() = "Decoder for Deep ASR model";
py::class_<Decoder>(m, "Decoder")
.def(py::init<std::string,
std::string,
std::string,
std::string,
size_t,
kaldi::BaseFloat>())
.def("decode_from_file",
(void (Decoder::*)(std::string, size_t)) & Decoder::decode_from_file,
"Decode for the probability matrices in specifier "
"and print the transcriptions.")
.def(
"decode",
(std::string (Decoder::*)(
std::string, const std::vector<std::vector<kaldi::BaseFloat>>&)) &
Decoder::decode,
"Decode one input probability matrix "
"and return the transcription.")
.def("decode_batch",
(std::vector<std::string> (Decoder::*)(
std::vector<std::string>,
const std::vector<std::vector<std::vector<kaldi::BaseFloat>>>&,
size_t num_processes)) &
Decoder::decode_batch,
"Decode one batch of probability matrices "
"and return the transcriptions.");
}
fluid/DeepASR/decoder/setup.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import glob
from distutils.core import setup, Extension
from distutils.sysconfig import get_config_vars
try:
kaldi_root = os.environ['KALDI_ROOT']
except:
raise ValueError("Enviroment variable 'KALDI_ROOT' is not defined. Please "
"install kaldi and export KALDI_ROOT=<kaldi's root dir> .")
args = [
'-std=c++11', '-fopenmp', '-Wno-sign-compare', '-Wno-unused-variable',
'-Wno-unused-local-typedefs', '-Wno-unused-but-set-variable',
'-Wno-deprecated-declarations', '-Wno-unused-function'
]
# remove warning about -Wstrict-prototypes
(opt, ) = get_config_vars('OPT')
os.environ['OPT'] = " ".join(flag for flag in opt.split()
if flag != '-Wstrict-prototypes')
os.environ['CC'] = 'g++'
LIBS = [
'fst', 'kaldi-base', 'kaldi-util', 'kaldi-matrix', 'kaldi-tree',
'kaldi-hmm', 'kaldi-fstext', 'kaldi-decoder', 'kaldi-lat'
]
LIB_DIRS = [
'tools/openfst/lib', 'src/base', 'src/matrix', 'src/util', 'src/tree',
'src/hmm', 'src/fstext', 'src/decoder', 'src/lat'
]
LIB_DIRS = [os.path.join(kaldi_root, path) for path in LIB_DIRS]
LIB_DIRS = [os.path.abspath(path) for path in LIB_DIRS]
ext_modules = [
Extension(
'post_latgen_faster_mapped',
['pybind.cc', 'post_latgen_faster_mapped.cc'],
include_dirs=[
'pybind11/include', '.', os.path.join(kaldi_root, 'src'),
os.path.join(kaldi_root, 'tools/openfst/src/include'), 'ThreadPool'
],
language='c++',
libraries=LIBS,
library_dirs=LIB_DIRS,
runtime_library_dirs=LIB_DIRS,
extra_compile_args=args, ),
]
setup(
name='post_latgen_faster_mapped',
version='0.1.0',
author='Paddle',
author_email='',
description='Decoder for Deep ASR model',
ext_modules=ext_modules, )
fluid/DeepASR/decoder/setup.sh 已删除 100644 → 0
浏览文件 @ 5ddb433e
set -e
if [ ! -d pybind11 ]; then
git clone https://github.com/pybind/pybind11.git
fi
if [ ! -d ThreadPool ]; then
git clone https://github.com/progschj/ThreadPool.git
echo -e "\n"
fi
python setup.py build_ext -i
fluid/DeepASR/examples/aishell/download_pretrained_model.sh 已删除 100644 → 0
浏览文件 @ 5ddb433e
url=http://deep-asr-data.gz.bcebos.com/aishell_pretrained_model.tar.gz
md5=7b51bde64e884f43901b7a3461ccbfa3
wget -c $url
echo "Checking md5 sum ..."
md5sum_tmp=`md5sum aishell_pretrained_model.tar.gz | cut -d ' ' -f1`
if [ $md5sum_tmp != $md5 ]; then
echo "Md5sum check failed, please remove and redownload "
"aishell_pretrained_model.tar.gz."
exit 1
fi
tar xvf aishell_pretrained_model.tar.gz
fluid/DeepASR/examples/aishell/infer_by_ckpt.sh 已删除 100644 → 0
浏览文件 @ 5ddb433e
decode_to_path=./decoding_result.txt
export CUDA_VISIBLE_DEVICES=0,1,2,3
python -u ../../infer_by_ckpt.py --batch_size 96 \
--checkpoint checkpoints/deep_asr.latest.checkpoint \
--infer_feature_lst data/test_feature.lst \
--mean_var data/global_mean_var \
--frame_dim 80 \
--class_num 3040 \
--num_threads 24 \
--beam_size 11 \
--decode_to_path $decode_to_path \
--trans_model aux/final.mdl \
--log_prior aux/logprior \
--vocabulary aux/graph/words.txt \
--graphs aux/graph/HCLG.fst \
--acoustic_scale 0.059 \
--parallel
fluid/DeepASR/examples/aishell/prepare_data.sh 已删除 100644 → 0
浏览文件 @ 5ddb433e
data_dir=~/.cache/paddle/dataset/speech/deep_asr_data/aishell
data_url='http://deep-asr-data.gz.bcebos.com/aishell_data.tar.gz'
lst_url='http://deep-asr-data.gz.bcebos.com/aishell_lst.tar.gz'
aux_url='http://deep-asr-data.gz.bcebos.com/aux.tar.gz'
md5=17669b8d63331c9326f4a9393d289bfb
aux_md5=50e3125eba1e3a2768a6f2e499cc1749
if [ ! -e $data_dir ]; then
mkdir -p $data_dir
fi
if [ ! -e $data_dir/aishell_data.tar.gz ]; then
echo "Download $data_dir/aishell_data.tar.gz ..."
wget -c -P $data_dir $data_url
else
echo "Skip downloading for $data_dir/aishell_data.tar.gz has already existed!"
fi
echo "Checking md5 sum ..."
md5sum_tmp=`md5sum $data_dir/aishell_data.tar.gz | cut -d ' ' -f1`
if [ $md5sum_tmp != $md5 ]; then
echo "Md5sum check failed, please remove and redownload "
"$data_dir/aishell_data.tar.gz"
exit 1
fi
echo "Untar aishell_data.tar.gz ..."
tar xzf $data_dir/aishell_data.tar.gz -C $data_dir
if [ ! -e data ]; then
mkdir data
fi
echo "Download and untar lst files ..."
wget -c -P data $lst_url
tar xvf data/aishell_lst.tar.gz -C data
ln -s $data_dir data/aishell
echo "Download and untar aux files ..."
wget -c $aux_url
tar xvf aux.tar.gz
fluid/DeepASR/examples/aishell/profile.sh 已删除 100644 → 0
浏览文件 @ 5ddb433e
export CUDA_VISIBLE_DEVICES=0
python -u ../../tools/profile.py --feature_lst data/train_feature.lst \
--label_lst data/train_label.lst \
--mean_var data/global_mean_var \
--frame_dim 80 \
--class_num 3040 \
--batch_size 16
fluid/DeepASR/examples/aishell/score_cer.sh 已删除 100644 → 0
浏览文件 @ 5ddb433e
ref_txt=aux/test.ref.txt
hyp_txt=decoding_result.txt
python ../../score_error_rate.py --error_rate_type cer --ref $ref_txt --hyp $hyp_txt
fluid/DeepASR/examples/aishell/train.sh 已删除 100644 → 0
浏览文件 @ 5ddb433e
export CUDA_VISIBLE_DEVICES=4,5,6,7
python -u ../../train.py --train_feature_lst data/train_feature.lst \
--train_label_lst data/train_label.lst \
--val_feature_lst data/val_feature.lst \
--val_label_lst data/val_label.lst \
--mean_var data/global_mean_var \
--checkpoints checkpoints \
--frame_dim 80 \
--class_num 3040 \
--print_per_batches 100 \
--infer_models '' \
--batch_size 16 \
--learning_rate 6.4e-5 \
--parallel
fluid/DeepASR/infer.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import os
import argparse
import paddle.fluid as fluid
import data_utils.augmentor.trans_mean_variance_norm as trans_mean_variance_norm
import data_utils.augmentor.trans_add_delta as trans_add_delta
import data_utils.augmentor.trans_splice as trans_splice
import data_utils.async_data_reader as reader
from data_utils.util import lodtensor_to_ndarray
from data_utils.util import split_infer_result
def parse_args():
parser = argparse.ArgumentParser("Inference for stacked LSTMP model.")
parser.add_argument(
'--batch_size',
type=int,
default=32,
help='The sequence number of a batch data. (default: %(default)d)')
parser.add_argument(
'--device',
type=str,
default='GPU',
choices=['CPU', 'GPU'],
help='The device type. (default: %(default)s)')
parser.add_argument(
'--mean_var',
type=str,
default='data/global_mean_var_search26kHr',
help="The path for feature's global mean and variance. "
"(default: %(default)s)")
parser.add_argument(
'--infer_feature_lst',
type=str,
default='data/infer_feature.lst',
help='The feature list path for inference. (default: %(default)s)')
parser.add_argument(
'--infer_label_lst',
type=str,
default='data/infer_label.lst',
help='The label list path for inference. (default: %(default)s)')
parser.add_argument(
'--infer_model_path',
type=str,
default='./infer_models/deep_asr.pass_0.infer.model/',
help='The directory for loading inference model. '
'(default: %(default)s)')
args = parser.parse_args()
return args
def print_arguments(args):
print('----------- Configuration Arguments -----------')
for arg, value in sorted(vars(args).iteritems()):
print('%s: %s' % (arg, value))
print('------------------------------------------------')
def infer(args):
""" Gets one batch of feature data and predicts labels for each sample.
"""
if not os.path.exists(args.infer_model_path):
raise IOError("Invalid inference model path!")
place = fluid.CUDAPlace(0) if args.device == 'GPU' else fluid.CPUPlace()
exe = fluid.Executor(place)
# load model
[infer_program, feed_dict,
fetch_targets] = fluid.io.load_inference_model(args.infer_model_path, exe)
ltrans = [
trans_add_delta.TransAddDelta(2, 2),
trans_mean_variance_norm.TransMeanVarianceNorm(args.mean_var),
trans_splice.TransSplice()
]
infer_data_reader = reader.AsyncDataReader(args.infer_feature_lst,
args.infer_label_lst)
infer_data_reader.set_transformers(ltrans)
feature_t = fluid.LoDTensor()
one_batch = infer_data_reader.batch_iterator(args.batch_size, 1).next()
(features, labels, lod) = one_batch
feature_t.set(features, place)
feature_t.set_lod([lod])
results = exe.run(infer_program,
feed={feed_dict[0]: feature_t},
fetch_list=fetch_targets,
return_numpy=False)
probs, lod = lodtensor_to_ndarray(results[0])
preds = probs.argmax(axis=1)
infer_batch = split_infer_result(preds, lod)
for index, sample in enumerate(infer_batch):
print("result %d: " % index, sample, '\n')
if __name__ == '__main__':
args = parse_args()
print_arguments(args)
infer(args)
fluid/DeepASR/infer_by_ckpt.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import sys
import os
import numpy as np
import argparse
import time
import paddle.fluid as fluid
import data_utils.augmentor.trans_mean_variance_norm as trans_mean_variance_norm
import data_utils.augmentor.trans_add_delta as trans_add_delta
import data_utils.augmentor.trans_splice as trans_splice
import data_utils.augmentor.trans_delay as trans_delay
import data_utils.async_data_reader as reader
from data_utils.util import lodtensor_to_ndarray, split_infer_result
from model_utils.model import stacked_lstmp_model
from decoder.post_latgen_faster_mapped import Decoder
from tools.error_rate import char_errors
def parse_args():
parser = argparse.ArgumentParser("Run inference by using checkpoint.")
parser.add_argument(
'--batch_size',
type=int,
default=32,
help='The sequence number of a batch data. (default: %(default)d)')
parser.add_argument(
'--beam_size',
type=int,
default=11,
help='The beam size for decoding. (default: %(default)d)')
parser.add_argument(
'--minimum_batch_size',
type=int,
default=1,
help='The minimum sequence number of a batch data. '
'(default: %(default)d)')
parser.add_argument(
'--frame_dim',
type=int,
default=80,
help='Frame dimension of feature data. (default: %(default)d)')
parser.add_argument(
'--stacked_num',
type=int,
default=5,
help='Number of lstmp layers to stack. (default: %(default)d)')
parser.add_argument(
'--proj_dim',
type=int,
default=512,
help='Project size of lstmp unit. (default: %(default)d)')
parser.add_argument(
'--hidden_dim',
type=int,
default=1024,
help='Hidden size of lstmp unit. (default: %(default)d)')
parser.add_argument(
'--class_num',
type=int,
default=1749,
help='Number of classes in label. (default: %(default)d)')
parser.add_argument(
'--num_threads',
type=int,
default=10,
help='The number of threads for decoding. (default: %(default)d)')
parser.add_argument(
'--device',
type=str,
default='GPU',
choices=['CPU', 'GPU'],
help='The device type. (default: %(default)s)')
parser.add_argument(
'--parallel', action='store_true', help='If set, run in parallel.')
parser.add_argument(
'--mean_var',
type=str,
default='data/global_mean_var',
help="The path for feature's global mean and variance. "
"(default: %(default)s)")
parser.add_argument(
'--infer_feature_lst',
type=str,
default='data/infer_feature.lst',
help='The feature list path for inference. (default: %(default)s)')
parser.add_argument(
'--checkpoint',
type=str,
default='./checkpoint',
help="The checkpoint path to init model. (default: %(default)s)")
parser.add_argument(
'--trans_model',
type=str,
default='./graph/trans_model',
help="The path to vocabulary. (default: %(default)s)")
parser.add_argument(
'--vocabulary',
type=str,
default='./graph/words.txt',
help="The path to vocabulary. (default: %(default)s)")
parser.add_argument(
'--graphs',
type=str,
default='./graph/TLG.fst',
help="The path to TLG graphs for decoding. (default: %(default)s)")
parser.add_argument(
'--log_prior',
type=str,
default="./logprior",
help="The log prior probs for training data. (default: %(default)s)")
parser.add_argument(
'--acoustic_scale',
type=float,
default=0.2,
help="Scaling factor for acoustic likelihoods. (default: %(default)f)")
parser.add_argument(
'--post_matrix_path',
type=str,
default=None,
help="The path to output post prob matrix. (default: %(default)s)")
parser.add_argument(
'--decode_to_path',
type=str,
default='./decoding_result.txt',
required=True,
help="The path to output the decoding result. (default: %(default)s)")
args = parser.parse_args()
return args
def print_arguments(args):
print('----------- Configuration Arguments -----------')
for arg, value in sorted(vars(args).iteritems()):
print('%s: %s' % (arg, value))
print('------------------------------------------------')
class PostMatrixWriter:
""" The writer for outputing the post probability matrix
"""
def __init__(self, to_path):
self._to_path = to_path
with open(self._to_path, "w") as post_matrix:
post_matrix.seek(0)
post_matrix.truncate()
def write(self, keys, probs):
with open(self._to_path, "a") as post_matrix:
if isinstance(keys, str):
keys, probs = [keys], [probs]
for key, prob in zip(keys, probs):
post_matrix.write(key + " [\n")
for i in range(prob.shape[0]):
for j in range(prob.shape[1]):
post_matrix.write(str(prob[i][j]) + " ")
post_matrix.write("\n")
post_matrix.write("]\n")
class DecodingResultWriter:
""" The writer for writing out decoding results
"""
def __init__(self, to_path):
self._to_path = to_path
with open(self._to_path, "w") as decoding_result:
decoding_result.seek(0)
decoding_result.truncate()
def write(self, results):
with open(self._to_path, "a") as decoding_result:
if isinstance(results, str):
decoding_result.write(results.encode("utf8") + "\n")
else:
for result in results:
decoding_result.write(result.encode("utf8") + "\n")
def infer_from_ckpt(args):
"""Inference by using checkpoint."""
if not os.path.exists(args.checkpoint):
raise IOError("Invalid checkpoint!")
prediction, avg_cost, accuracy = stacked_lstmp_model(
frame_dim=args.frame_dim,
hidden_dim=args.hidden_dim,
proj_dim=args.proj_dim,
stacked_num=args.stacked_num,
class_num=args.class_num,
parallel=args.parallel)
infer_program = fluid.default_main_program().clone()
# optimizer, placeholder
optimizer = fluid.optimizer.Adam(
learning_rate=fluid.layers.exponential_decay(
learning_rate=0.0001,
decay_steps=1879,
decay_rate=1 / 1.2,
staircase=True))
optimizer.minimize(avg_cost)
place = fluid.CPUPlace() if args.device == 'CPU' else fluid.CUDAPlace(0)
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())
# load checkpoint.
fluid.io.load_persistables(exe, args.checkpoint)
# init decoder
decoder = Decoder(args.trans_model, args.vocabulary, args.graphs,
args.log_prior, args.beam_size, args.acoustic_scale)
ltrans = [
trans_add_delta.TransAddDelta(2, 2),
trans_mean_variance_norm.TransMeanVarianceNorm(args.mean_var),
trans_splice.TransSplice(5, 5), trans_delay.TransDelay(5)
]
feature_t = fluid.LoDTensor()
label_t = fluid.LoDTensor()
# infer data reader
infer_data_reader = reader.AsyncDataReader(
args.infer_feature_lst, drop_frame_len=-1, split_sentence_threshold=-1)
infer_data_reader.set_transformers(ltrans)
decoding_result_writer = DecodingResultWriter(args.decode_to_path)
post_matrix_writer = None if args.post_matrix_path is None \
else PostMatrixWriter(args.post_matrix_path)
for batch_id, batch_data in enumerate(
infer_data_reader.batch_iterator(args.batch_size,
args.minimum_batch_size)):
# load_data
(features, labels, lod, name_lst) = batch_data
features = np.reshape(features, (-1, 11, 3, args.frame_dim))
features = np.transpose(features, (0, 2, 1, 3))
feature_t.set(features, place)
feature_t.set_lod([lod])
label_t.set(labels, place)
label_t.set_lod([lod])
results = exe.run(infer_program,
feed={"feature": feature_t,
"label": label_t},
fetch_list=[prediction, avg_cost, accuracy],
return_numpy=False)
probs, lod = lodtensor_to_ndarray(results[0])
infer_batch = split_infer_result(probs, lod)
print("Decoding batch %d ..." % batch_id)
decoded = decoder.decode_batch(name_lst, infer_batch, args.num_threads)
decoding_result_writer.write(decoded)
if args.post_matrix_path is not None:
post_matrix_writer.write(name_lst, infer_batch)
if __name__ == '__main__':
args = parse_args()
print_arguments(args)
infer_from_ckpt(args)
fluid/DeepASR/model_utils/model.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import paddle.fluid as fluid
def stacked_lstmp_model(feature,
label,
hidden_dim,
proj_dim,
stacked_num,
class_num,
parallel=False,
is_train=True):
"""
The model for DeepASR. The main structure is composed of stacked
identical LSTMP (LSTM with recurrent projection) layers.
When running in training and validation phase, the feeding dictionary
is {'feature', 'label'}, fed by the LodTensor for feature data and
label data respectively. And in inference, only `feature` is needed.
Args:
frame_dim(int): The frame dimension of feature data.
hidden_dim(int): The hidden state's dimension of the LSTMP layer.
proj_dim(int): The projection size of the LSTMP layer.
stacked_num(int): The number of stacked LSTMP layers.
parallel(bool): Run in parallel or not, default `False`.
is_train(bool): Run in training phase or not, default `True`.
class_dim(int): The number of output classes.
"""
conv1 = fluid.layers.conv2d(
input=feature,
num_filters=32,
filter_size=3,
stride=1,
padding=1,
bias_attr=True,
act="relu")
pool1 = fluid.layers.pool2d(
conv1, pool_size=3, pool_type="max", pool_stride=2, pool_padding=0)
stack_input = pool1
for i in range(stacked_num):
fc = fluid.layers.fc(input=stack_input,
size=hidden_dim * 4,
bias_attr=None)
proj, cell = fluid.layers.dynamic_lstmp(
input=fc,
size=hidden_dim * 4,
proj_size=proj_dim,
bias_attr=True,
use_peepholes=True,
is_reverse=False,
cell_activation="tanh",
proj_activation="tanh")
bn = fluid.layers.batch_norm(
input=proj,
is_test=not is_train,
momentum=0.9,
epsilon=1e-05,
data_layout='NCHW')
stack_input = bn
prediction = fluid.layers.fc(input=stack_input,
size=class_num,
act='softmax')
cost = fluid.layers.cross_entropy(input=prediction, label=label)
avg_cost = fluid.layers.mean(x=cost)
acc = fluid.layers.accuracy(input=prediction, label=label)
return prediction, avg_cost, acc
fluid/DeepASR/score_error_rate.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import argparse
from tools.error_rate import char_errors, word_errors
def parse_args():
parser = argparse.ArgumentParser(
"Score word/character error rate (WER/CER) "
"for decoding result.")
parser.add_argument(
'--error_rate_type',
type=str,
default='cer',
choices=['cer', 'wer'],
help="Error rate type. (default: %(default)s)")
parser.add_argument(
'--special_tokens',
type=str,
default='<SPOKEN_NOISE>',
help="Special tokens in scoring CER, seperated by space. "
"They shouldn't be splitted and should be treated as one special "
"character. Example: '<SPOKEN_NOISE> <bos> <eos>' "
"(default: %(default)s)")
parser.add_argument(
'--ref', type=str, required=True, help="The ground truth text.")
parser.add_argument(
'--hyp', type=str, required=True, help="The decoding result text.")
args = parser.parse_args()
return args
if __name__ == '__main__':
args = parse_args()
ref_dict = {}
sum_errors, sum_ref_len = 0.0, 0
sent_cnt, not_in_ref_cnt = 0, 0
special_tokens = args.special_tokens.split(" ")
with open(args.ref, "r") as ref_txt:
line = ref_txt.readline()
while line:
del_pos = line.find(" ")
key, sent = line[0:del_pos], line[del_pos + 1:-1].strip()
ref_dict[key] = sent
line = ref_txt.readline()
with open(args.hyp, "r") as hyp_txt:
line = hyp_txt.readline()
while line:
del_pos = line.find(" ")
key, sent = line[0:del_pos], line[del_pos + 1:-1].strip()
sent_cnt += 1
line = hyp_txt.readline()
if key not in ref_dict:
not_in_ref_cnt += 1
continue
if args.error_rate_type == 'cer':
for sp_tok in special_tokens:
sent = sent.replace(sp_tok, '\0')
errors, ref_len = char_errors(
ref_dict[key].decode("utf8"),
sent.decode("utf8"),
remove_space=True)
else:
errors, ref_len = word_errors(ref_dict[key].decode("utf8"),
sent.decode("utf8"))
sum_errors += errors
sum_ref_len += ref_len
print("Error rate[%s] = %f (%d/%d)," %
(args.error_rate_type, sum_errors / sum_ref_len, int(sum_errors),
sum_ref_len))
print("total %d sentences in hyp, %d not presented in ref." %
(sent_cnt, not_in_ref_cnt))
fluid/DeepASR/tools/_init_paths.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
"""Add the parent directory to $PYTHONPATH"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import os.path
import sys
def add_path(path):
if path not in sys.path:
sys.path.insert(0, path)
this_dir = os.path.dirname(__file__)
# Add project path to PYTHONPATH
proj_path = os.path.join(this_dir, '..')
add_path(proj_path)
fluid/DeepASR/tools/error_rate.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
# -*- coding: utf-8 -*-
"""This module provides functions to calculate error rate in different level.
e.g. wer for word-level, cer for char-level.
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import numpy as np
def _levenshtein_distance(ref, hyp):
"""Levenshtein distance is a string metric for measuring the difference
between two sequences. Informally, the levenshtein disctance is defined as
the minimum number of single-character edits (substitutions, insertions or
deletions) required to change one word into the other. We can naturally
extend the edits to word level when calculate levenshtein disctance for
two sentences.
"""
m = len(ref)
n = len(hyp)
# special case
if ref == hyp:
return 0
if m == 0:
return n
if n == 0:
return m
if m < n:
ref, hyp = hyp, ref
m, n = n, m
# use O(min(m, n)) space
distance = np.zeros((2, n + 1), dtype=np.int32)
# initialize distance matrix
for j in xrange(n + 1):
distance[0][j] = j
# calculate levenshtein distance
for i in xrange(1, m + 1):
prev_row_idx = (i - 1) % 2
cur_row_idx = i % 2
distance[cur_row_idx][0] = i
for j in xrange(1, n + 1):
if ref[i - 1] == hyp[j - 1]:
distance[cur_row_idx][j] = distance[prev_row_idx][j - 1]
else:
s_num = distance[prev_row_idx][j - 1] + 1
i_num = distance[cur_row_idx][j - 1] + 1
d_num = distance[prev_row_idx][j] + 1
distance[cur_row_idx][j] = min(s_num, i_num, d_num)
return distance[m % 2][n]
def word_errors(reference, hypothesis, ignore_case=False, delimiter=' '):
"""Compute the levenshtein distance between reference sequence and
hypothesis sequence in word-level.
:param reference: The reference sentence.
:type reference: basestring
:param hypothesis: The hypothesis sentence.
:type hypothesis: basestring
:param ignore_case: Whether case-sensitive or not.
:type ignore_case: bool
:param delimiter: Delimiter of input sentences.
:type delimiter: char
:return: Levenshtein distance and word number of reference sentence.
:rtype: list
"""
if ignore_case == True:
reference = reference.lower()
hypothesis = hypothesis.lower()
ref_words = filter(None, reference.split(delimiter))
hyp_words = filter(None, hypothesis.split(delimiter))
edit_distance = _levenshtein_distance(ref_words, hyp_words)
return float(edit_distance), len(ref_words)
def char_errors(reference, hypothesis, ignore_case=False, remove_space=False):
"""Compute the levenshtein distance between reference sequence and
hypothesis sequence in char-level.
:param reference: The reference sentence.
:type reference: basestring
:param hypothesis: The hypothesis sentence.
:type hypothesis: basestring
:param ignore_case: Whether case-sensitive or not.
:type ignore_case: bool
:param remove_space: Whether remove internal space characters
:type remove_space: bool
:return: Levenshtein distance and length of reference sentence.
:rtype: list
"""
if ignore_case == True:
reference = reference.lower()
hypothesis = hypothesis.lower()
join_char = ' '
if remove_space == True:
join_char = ''
reference = join_char.join(filter(None, reference.split(' ')))
hypothesis = join_char.join(filter(None, hypothesis.split(' ')))
edit_distance = _levenshtein_distance(reference, hypothesis)
return float(edit_distance), len(reference)
def wer(reference, hypothesis, ignore_case=False, delimiter=' '):
"""Calculate word error rate (WER). WER compares reference text and
hypothesis text in word-level. WER is defined as:
.. math::
WER = (Sw + Dw + Iw) / Nw
where
.. code-block:: text
Sw is the number of words subsituted,
Dw is the number of words deleted,
Iw is the number of words inserted,
Nw is the number of words in the reference
We can use levenshtein distance to calculate WER. Please draw an attention
that empty items will be removed when splitting sentences by delimiter.
:param reference: The reference sentence.
:type reference: basestring
:param hypothesis: The hypothesis sentence.
:type hypothesis: basestring
:param ignore_case: Whether case-sensitive or not.
:type ignore_case: bool
:param delimiter: Delimiter of input sentences.
:type delimiter: char
:return: Word error rate.
:rtype: float
:raises ValueError: If word number of reference is zero.
"""
edit_distance, ref_len = word_errors(reference, hypothesis, ignore_case,
delimiter)
if ref_len == 0:
raise ValueError("Reference's word number should be greater than 0.")
wer = float(edit_distance) / ref_len
return wer
def cer(reference, hypothesis, ignore_case=False, remove_space=False):
"""Calculate charactor error rate (CER). CER compares reference text and
hypothesis text in char-level. CER is defined as:
.. math::
CER = (Sc + Dc + Ic) / Nc
where
.. code-block:: text
Sc is the number of characters substituted,
Dc is the number of characters deleted,
Ic is the number of characters inserted
Nc is the number of characters in the reference
We can use levenshtein distance to calculate CER. Chinese input should be
encoded to unicode. Please draw an attention that the leading and tailing
space characters will be truncated and multiple consecutive space
characters in a sentence will be replaced by one space character.
:param reference: The reference sentence.
:type reference: basestring
:param hypothesis: The hypothesis sentence.
:type hypothesis: basestring
:param ignore_case: Whether case-sensitive or not.
:type ignore_case: bool
:param remove_space: Whether remove internal space characters
:type remove_space: bool
:return: Character error rate.
:rtype: float
:raises ValueError: If the reference length is zero.
"""
edit_distance, ref_len = char_errors(reference, hypothesis, ignore_case,
remove_space)
if ref_len == 0:
raise ValueError("Length of reference should be greater than 0.")
cer = float(edit_distance) / ref_len
return cer
fluid/DeepASR/tools/profile.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import sys
import numpy as np
import argparse
import time
import paddle.fluid as fluid
import paddle.fluid.profiler as profiler
import _init_paths
import data_utils.augmentor.trans_mean_variance_norm as trans_mean_variance_norm
import data_utils.augmentor.trans_add_delta as trans_add_delta
import data_utils.augmentor.trans_splice as trans_splice
import data_utils.augmentor.trans_delay as trans_delay
import data_utils.async_data_reader as reader
from model_utils.model import stacked_lstmp_model
from data_utils.util import lodtensor_to_ndarray
def parse_args():
parser = argparse.ArgumentParser("Profiling for the stacked LSTMP model.")
parser.add_argument(
'--batch_size',
type=int,
default=32,
help='The sequence number of a batch data. (default: %(default)d)')
parser.add_argument(
'--minimum_batch_size',
type=int,
default=1,
help='The minimum sequence number of a batch data. '
'(default: %(default)d)')
parser.add_argument(
'--frame_dim',
type=int,
default=120 * 11,
help='Frame dimension of feature data. (default: %(default)d)')
parser.add_argument(
'--stacked_num',
type=int,
default=5,
help='Number of lstmp layers to stack. (default: %(default)d)')
parser.add_argument(
'--proj_dim',
type=int,
default=512,
help='Project size of lstmp unit. (default: %(default)d)')
parser.add_argument(
'--hidden_dim',
type=int,
default=1024,
help='Hidden size of lstmp unit. (default: %(default)d)')
parser.add_argument(
'--class_num',
type=int,
default=1749,
help='Number of classes in label. (default: %(default)d)')
parser.add_argument(
'--learning_rate',
type=float,
default=0.00016,
help='Learning rate used to train. (default: %(default)f)')
parser.add_argument(
'--device',
type=str,
default='GPU',
choices=['CPU', 'GPU'],
help='The device type. (default: %(default)s)')
parser.add_argument(
'--parallel', action='store_true', help='If set, run in parallel.')
parser.add_argument(
'--mean_var',
type=str,
default='data/global_mean_var_search26kHr',
help='mean var path')
parser.add_argument(
'--feature_lst',
type=str,
default='data/feature.lst',
help='feature list path.')
parser.add_argument(
'--label_lst',
type=str,
default='data/label.lst',
help='label list path.')
parser.add_argument(
'--max_batch_num',
type=int,
default=11,
help='Maximum number of batches for profiling. (default: %(default)d)')
parser.add_argument(
'--first_batches_to_skip',
type=int,
default=1,
help='Number of first batches to skip for profiling. '
'(default: %(default)d)')
parser.add_argument(
'--print_train_acc',
action='store_true',
help='If set, output training accuray.')
parser.add_argument(
'--sorted_key',
type=str,
default='total',
choices=['None', 'total', 'calls', 'min', 'max', 'ave'],
help='Different types of time to sort the profiling report. '
'(default: %(default)s)')
args = parser.parse_args()
return args
def print_arguments(args):
print('----------- Configuration Arguments -----------')
for arg, value in sorted(vars(args).iteritems()):
print('%s: %s' % (arg, value))
print('------------------------------------------------')
def profile(args):
"""profile the training process.
"""
if not args.first_batches_to_skip < args.max_batch_num:
raise ValueError("arg 'first_batches_to_skip' must be smaller than "
"'max_batch_num'.")
if not args.first_batches_to_skip >= 0:
raise ValueError(
"arg 'first_batches_to_skip' must not be smaller than 0.")
_, avg_cost, accuracy = stacked_lstmp_model(
frame_dim=args.frame_dim,
hidden_dim=args.hidden_dim,
proj_dim=args.proj_dim,
stacked_num=args.stacked_num,
class_num=args.class_num,
parallel=args.parallel)
optimizer = fluid.optimizer.Adam(
learning_rate=fluid.layers.exponential_decay(
learning_rate=args.learning_rate,
decay_steps=1879,
decay_rate=1 / 1.2,
staircase=True))
optimizer.minimize(avg_cost)
place = fluid.CPUPlace() if args.device == 'CPU' else fluid.CUDAPlace(0)
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())
ltrans = [
trans_add_delta.TransAddDelta(2, 2),
trans_mean_variance_norm.TransMeanVarianceNorm(args.mean_var),
trans_splice.TransSplice(5, 5), trans_delay.TransDelay(5)
]
data_reader = reader.AsyncDataReader(
args.feature_lst, args.label_lst, -1, split_sentence_threshold=1024)
data_reader.set_transformers(ltrans)
feature_t = fluid.LoDTensor()
label_t = fluid.LoDTensor()
sorted_key = None if args.sorted_key is 'None' else args.sorted_key
with profiler.profiler(args.device, sorted_key) as prof:
frames_seen, start_time = 0, 0.0
for batch_id, batch_data in enumerate(
data_reader.batch_iterator(args.batch_size,
args.minimum_batch_size)):
if batch_id >= args.max_batch_num:
break
if args.first_batches_to_skip == batch_id:
profiler.reset_profiler()
start_time = time.time()
frames_seen = 0
# load_data
(features, labels, lod, _) = batch_data
features = np.reshape(features, (-1, 11, 3, args.frame_dim))
features = np.transpose(features, (0, 2, 1, 3))
feature_t.set(features, place)
feature_t.set_lod([lod])
label_t.set(labels, place)
label_t.set_lod([lod])
frames_seen += lod[-1]
outs = exe.run(fluid.default_main_program(),
feed={"feature": feature_t,
"label": label_t},
fetch_list=[avg_cost, accuracy]
if args.print_train_acc else [],
return_numpy=False)
if args.print_train_acc:
print("Batch %d acc: %f" %
(batch_id, lodtensor_to_ndarray(outs[1])[0]))
else:
sys.stdout.write('.')
sys.stdout.flush()
time_consumed = time.time() - start_time
frames_per_sec = frames_seen / time_consumed
print("\nTime consumed: %f s, performance: %f frames/s." %
(time_consumed, frames_per_sec))
if __name__ == '__main__':
args = parse_args()
print_arguments(args)
profile(args)
fluid/DeepASR/train.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import sys
import os
import numpy as np
import argparse
import time
import paddle.fluid as fluid
import data_utils.augmentor.trans_mean_variance_norm as trans_mean_variance_norm
import data_utils.augmentor.trans_add_delta as trans_add_delta
import data_utils.augmentor.trans_splice as trans_splice
import data_utils.augmentor.trans_delay as trans_delay
import data_utils.async_data_reader as reader
from model_utils.model import stacked_lstmp_model
def parse_args():
parser = argparse.ArgumentParser("Training for stacked LSTMP model.")
parser.add_argument(
'--batch_size',
type=int,
default=32,
help='The sequence number of a batch data. Batch size per GPU. (default: %(default)d)'
)
parser.add_argument(
'--minimum_batch_size',
type=int,
default=1,
help='The minimum sequence number of a batch data. '
'(default: %(default)d)')
parser.add_argument(
'--frame_dim',
type=int,
default=80,
help='Frame dimension of feature data. (default: %(default)d)')
parser.add_argument(
'--stacked_num',
type=int,
default=5,
help='Number of lstmp layers to stack. (default: %(default)d)')
parser.add_argument(
'--proj_dim',
type=int,
default=512,
help='Project size of lstmp unit. (default: %(default)d)')
parser.add_argument(
'--hidden_dim',
type=int,
default=1024,
help='Hidden size of lstmp unit. (default: %(default)d)')
parser.add_argument(
'--class_num',
type=int,
default=3040,
help='Number of classes in label. (default: %(default)d)')
parser.add_argument(
'--pass_num',
type=int,
default=100,
help='Epoch number to train. (default: %(default)d)')
parser.add_argument(
'--print_per_batches',
type=int,
default=100,
help='Interval to print training accuracy. (default: %(default)d)')
parser.add_argument(
'--learning_rate',
type=float,
default=0.00016,
help='Learning rate used to train. (default: %(default)f)')
parser.add_argument(
'--device',
type=str,
default='GPU',
choices=['CPU', 'GPU'],
help='The device type. (default: %(default)s)')
parser.add_argument(
'--parallel', action='store_true', help='If set, run in parallel.')
parser.add_argument(
'--mean_var',
type=str,
default='data/global_mean_var_search26kHr',
help="The path for feature's global mean and variance. "
"(default: %(default)s)")
parser.add_argument(
'--train_feature_lst',
type=str,
default='data/feature.lst',
help='The feature list path for training. (default: %(default)s)')
parser.add_argument(
'--train_label_lst',
type=str,
default='data/label.lst',
help='The label list path for training. (default: %(default)s)')
parser.add_argument(
'--val_feature_lst',
type=str,
default='data/val_feature.lst',
help='The feature list path for validation. (default: %(default)s)')
parser.add_argument(
'--val_label_lst',
type=str,
default='data/val_label.lst',
help='The label list path for validation. (default: %(default)s)')
parser.add_argument(
'--init_model_path',
type=str,
default=None,
help="The model (checkpoint) path which the training resumes from. "
"If None, train the model from scratch. (default: %(default)s)")
parser.add_argument(
'--checkpoints',
type=str,
default='./checkpoints',
help="The directory for saving checkpoints. Do not save checkpoints "
"if set to ''. (default: %(default)s)")
parser.add_argument(
'--infer_models',
type=str,
default='./infer_models',
help="The directory for saving inference models. Do not save inference "
"models if set to ''. (default: %(default)s)")
args = parser.parse_args()
return args
def print_arguments(args):
print('----------- Configuration Arguments -----------')
for arg, value in sorted(vars(args).iteritems()):
print('%s: %s' % (arg, value))
print('------------------------------------------------')
def train(args):
"""train in loop.
"""
# paths check
if args.init_model_path is not None and \
not os.path.exists(args.init_model_path):
raise IOError("Invalid initial model path!")
if args.checkpoints != '' and not os.path.exists(args.checkpoints):
os.mkdir(args.checkpoints)
if args.infer_models != '' and not os.path.exists(args.infer_models):
os.mkdir(args.infer_models)
train_program = fluid.Program()
train_startup = fluid.Program()
with fluid.program_guard(train_program, train_startup):
with fluid.unique_name.guard():
py_train_reader = fluid.layers.py_reader(
capacity=10,
shapes=([-1, 3, 11, args.frame_dim], [-1, 1]),
dtypes=['float32', 'int64'],
lod_levels=[1, 1],
name='train_reader')
feature, label = fluid.layers.read_file(py_train_reader)
prediction, avg_cost, accuracy = stacked_lstmp_model(
feature=feature,
label=label,
hidden_dim=args.hidden_dim,
proj_dim=args.proj_dim,
stacked_num=args.stacked_num,
class_num=args.class_num)
# optimizer = fluid.optimizer.Momentum(learning_rate=args.learning_rate, momentum=0.9)
optimizer = fluid.optimizer.Adam(
learning_rate=fluid.layers.exponential_decay(
learning_rate=args.learning_rate,
decay_steps=1879,
decay_rate=1 / 1.2,
staircase=True))
optimizer.minimize(avg_cost)
fluid.memory_optimize(train_program)
test_program = fluid.Program()
test_startup = fluid.Program()
with fluid.program_guard(test_program, test_startup):
with fluid.unique_name.guard():
py_test_reader = fluid.layers.py_reader(
capacity=10,
shapes=([-1, 3, 11, args.frame_dim], [-1, 1]),
dtypes=['float32', 'int64'],
lod_levels=[1, 1],
name='test_reader')
feature, label = fluid.layers.read_file(py_test_reader)
prediction, avg_cost, accuracy = stacked_lstmp_model(
feature=feature,
label=label,
hidden_dim=args.hidden_dim,
proj_dim=args.proj_dim,
stacked_num=args.stacked_num,
class_num=args.class_num)
test_program = test_program.clone(for_test=True)
place = fluid.CPUPlace() if args.device == 'CPU' else fluid.CUDAPlace(0)
exe = fluid.Executor(place)
exe.run(train_startup)
exe.run(test_startup)
if args.parallel:
exec_strategy = fluid.ExecutionStrategy()
exec_strategy.num_iteration_per_drop_scope = 10
train_exe = fluid.ParallelExecutor(
use_cuda=(args.device == 'GPU'),
loss_name=avg_cost.name,
exec_strategy=exec_strategy,
main_program=train_program)
test_exe = fluid.ParallelExecutor(
use_cuda=(args.device == 'GPU'),
main_program=test_program,
exec_strategy=exec_strategy,
share_vars_from=train_exe)
# resume training if initial model provided.
if args.init_model_path is not None:
fluid.io.load_persistables(exe, args.init_model_path)
ltrans = [
trans_add_delta.TransAddDelta(2, 2),
trans_mean_variance_norm.TransMeanVarianceNorm(args.mean_var),
trans_splice.TransSplice(5, 5), trans_delay.TransDelay(5)
]
# bind train_reader
train_data_reader = reader.AsyncDataReader(
args.train_feature_lst,
args.train_label_lst,
-1,
split_sentence_threshold=1024)
train_data_reader.set_transformers(ltrans)
def train_data_provider():
for data in train_data_reader.batch_iterator(args.batch_size,
args.minimum_batch_size):
yield batch_data_to_lod_tensors(args, data, fluid.CPUPlace())
py_train_reader.decorate_tensor_provider(train_data_provider)
if (os.path.exists(args.val_feature_lst) and
os.path.exists(args.val_label_lst)):
# test data reader
test_data_reader = reader.AsyncDataReader(
args.val_feature_lst,
args.val_label_lst,
-1,
split_sentence_threshold=1024)
test_data_reader.set_transformers(ltrans)
def test_data_provider():
for data in test_data_reader.batch_iterator(
args.batch_size, args.minimum_batch_size):
yield batch_data_to_lod_tensors(args, data, fluid.CPUPlace())
py_test_reader.decorate_tensor_provider(test_data_provider)
# validation
def test(exe):
# If test data not found, return invalid cost and accuracy
if not (os.path.exists(args.val_feature_lst) and
os.path.exists(args.val_label_lst)):
return -1.0, -1.0
batch_id = 0
test_costs = []
test_accs = []
while True:
if batch_id == 0:
py_test_reader.start()
try:
if args.parallel:
cost, acc = exe.run(
fetch_list=[avg_cost.name, accuracy.name],
return_numpy=False)
else:
cost, acc = exe.run(program=test_program,
fetch_list=[avg_cost, accuracy],
return_numpy=False)
sys.stdout.write('.')
sys.stdout.flush()
test_costs.append(np.array(cost)[0])
test_accs.append(np.array(acc)[0])
batch_id += 1
except fluid.core.EOFException:
py_test_reader.reset()
break
return np.mean(test_costs), np.mean(test_accs)
# train
for pass_id in xrange(args.pass_num):
pass_start_time = time.time()
batch_id = 0
while True:
if batch_id == 0:
py_train_reader.start()
to_print = batch_id > 0 and (batch_id % args.print_per_batches == 0)
try:
if args.parallel:
outs = train_exe.run(
fetch_list=[avg_cost.name, accuracy.name]
if to_print else [],
return_numpy=False)
else:
outs = exe.run(program=train_program,
fetch_list=[avg_cost, accuracy]
if to_print else [],
return_numpy=False)
except fluid.core.EOFException:
py_train_reader.reset()
break
if to_print:
if args.parallel:
print("\nBatch %d, train cost: %f, train acc: %f" %
(batch_id, np.mean(outs[0]), np.mean(outs[1])))
else:
print("\nBatch %d, train cost: %f, train acc: %f" % (
batch_id, np.array(outs[0])[0], np.array(outs[1])[0]))
# save the latest checkpoint
if args.checkpoints != '':
model_path = os.path.join(args.checkpoints,
"deep_asr.latest.checkpoint")
fluid.io.save_persistables(exe, model_path, train_program)
else:
sys.stdout.write('.')
sys.stdout.flush()
batch_id += 1
# run test
val_cost, val_acc = test(test_exe if args.parallel else exe)
# save checkpoint per pass
if args.checkpoints != '':
model_path = os.path.join(
args.checkpoints,
"deep_asr.pass_" + str(pass_id) + ".checkpoint")
fluid.io.save_persistables(exe, model_path, train_program)
# save inference model
if args.infer_models != '':
model_path = os.path.join(
args.infer_models,
"deep_asr.pass_" + str(pass_id) + ".infer.model")
fluid.io.save_inference_model(model_path, ["feature"],
[prediction], exe, train_program)
# cal pass time
pass_end_time = time.time()
time_consumed = pass_end_time - pass_start_time
# print info at pass end
print("\nPass %d, time consumed: %f s, val cost: %f, val acc: %f\n" %
(pass_id, time_consumed, val_cost, val_acc))
def batch_data_to_lod_tensors(args, batch_data, place):
features, labels, lod, name_lst = batch_data
features = np.reshape(features, (-1, 11, 3, args.frame_dim))
features = np.transpose(features, (0, 2, 1, 3))
feature_t = fluid.LoDTensor()
label_t = fluid.LoDTensor()
feature_t.set(features, place)
feature_t.set_lod([lod])
label_t.set(labels, place)
label_t.set_lod([lod])
return feature_t, label_t
if __name__ == '__main__':
args = parse_args()
print_arguments(args)
train(args)
fluid/DeepQNetwork/DQN_agent.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
#-*- coding: utf-8 -*-
import math
import numpy as np
import paddle.fluid as fluid
from paddle.fluid.param_attr import ParamAttr
from tqdm import tqdm
class DQNModel(object):
def __init__(self, state_dim, action_dim, gamma, hist_len, use_cuda=False):
self.img_height = state_dim[0]
self.img_width = state_dim[1]
self.action_dim = action_dim
self.gamma = gamma
self.exploration = 1.1
self.update_target_steps = 10000 // 4
self.hist_len = hist_len
self.use_cuda = use_cuda
self.global_step = 0
self._build_net()
def _get_inputs(self):
return fluid.layers.data(
name='state',
shape=[self.hist_len, self.img_height, self.img_width],
dtype='float32'), \
fluid.layers.data(
name='action', shape=[1], dtype='int32'), \
fluid.layers.data(
name='reward', shape=[], dtype='float32'), \
fluid.layers.data(
name='next_s',
shape=[self.hist_len, self.img_height, self.img_width],
dtype='float32'), \
fluid.layers.data(
name='isOver', shape=[], dtype='bool')
def _build_net(self):
self.predict_program = fluid.Program()
self.train_program = fluid.Program()
self._sync_program = fluid.Program()
with fluid.program_guard(self.predict_program):
state, action, reward, next_s, isOver = self._get_inputs()
self.pred_value = self.get_DQN_prediction(state)
with fluid.program_guard(self.train_program):
state, action, reward, next_s, isOver = self._get_inputs()
pred_value = self.get_DQN_prediction(state)
reward = fluid.layers.clip(reward, min=-1.0, max=1.0)
action_onehot = fluid.layers.one_hot(action, self.action_dim)
action_onehot = fluid.layers.cast(action_onehot, dtype='float32')
pred_action_value = fluid.layers.reduce_sum(
fluid.layers.elementwise_mul(action_onehot, pred_value), dim=1)
targetQ_predict_value = self.get_DQN_prediction(next_s, target=True)
best_v = fluid.layers.reduce_max(targetQ_predict_value, dim=1)
best_v.stop_gradient = True
target = reward + (1.0 - fluid.layers.cast(
isOver, dtype='float32')) * self.gamma * best_v
cost = fluid.layers.square_error_cost(pred_action_value, target)
cost = fluid.layers.reduce_mean(cost)
optimizer = fluid.optimizer.Adam(1e-3 * 0.5, epsilon=1e-3)
optimizer.minimize(cost)
vars = list(self.train_program.list_vars())
policy_vars = list(filter(
lambda x: 'GRAD' not in x.name and 'policy' in x.name, vars))
target_vars = list(filter(
lambda x: 'GRAD' not in x.name and 'target' in x.name, vars))
policy_vars.sort(key=lambda x: x.name)
target_vars.sort(key=lambda x: x.name)
with fluid.program_guard(self._sync_program):
sync_ops = []
for i, var in enumerate(policy_vars):
sync_op = fluid.layers.assign(policy_vars[i], target_vars[i])
sync_ops.append(sync_op)
# fluid exe
place = fluid.CUDAPlace(0) if self.use_cuda else fluid.CPUPlace()
self.exe = fluid.Executor(place)
self.exe.run(fluid.default_startup_program())
def get_DQN_prediction(self, image, target=False):
image = image / 255.0
variable_field = 'target' if target else 'policy'
conv1 = fluid.layers.conv2d(
input=image,
num_filters=32,
filter_size=5,
stride=1,
padding=2,
act='relu',
param_attr=ParamAttr(name='{}_conv1'.format(variable_field)),
bias_attr=ParamAttr(name='{}_conv1_b'.format(variable_field)))
max_pool1 = fluid.layers.pool2d(
input=conv1, pool_size=2, pool_stride=2, pool_type='max')
conv2 = fluid.layers.conv2d(
input=max_pool1,
num_filters=32,
filter_size=5,
stride=1,
padding=2,
act='relu',
param_attr=ParamAttr(name='{}_conv2'.format(variable_field)),
bias_attr=ParamAttr(name='{}_conv2_b'.format(variable_field)))
max_pool2 = fluid.layers.pool2d(
input=conv2, pool_size=2, pool_stride=2, pool_type='max')
conv3 = fluid.layers.conv2d(
input=max_pool2,
num_filters=64,
filter_size=4,
stride=1,
padding=1,
act='relu',
param_attr=ParamAttr(name='{}_conv3'.format(variable_field)),
bias_attr=ParamAttr(name='{}_conv3_b'.format(variable_field)))
max_pool3 = fluid.layers.pool2d(
input=conv3, pool_size=2, pool_stride=2, pool_type='max')
conv4 = fluid.layers.conv2d(
input=max_pool3,
num_filters=64,
filter_size=3,
stride=1,
padding=1,
act='relu',
param_attr=ParamAttr(name='{}_conv4'.format(variable_field)),
bias_attr=ParamAttr(name='{}_conv4_b'.format(variable_field)))
flatten = fluid.layers.flatten(conv4, axis=1)
out = fluid.layers.fc(
input=flatten,
size=self.action_dim,
param_attr=ParamAttr(name='{}_fc1'.format(variable_field)),
bias_attr=ParamAttr(name='{}_fc1_b'.format(variable_field)))
return out
def act(self, state, train_or_test):
sample = np.random.random()
if train_or_test == 'train' and sample < self.exploration:
act = np.random.randint(self.action_dim)
else:
if np.random.random() < 0.01:
act = np.random.randint(self.action_dim)
else:
state = np.expand_dims(state, axis=0)
pred_Q = self.exe.run(self.predict_program,
feed={'state': state.astype('float32')},
fetch_list=[self.pred_value])[0]
pred_Q = np.squeeze(pred_Q, axis=0)
act = np.argmax(pred_Q)
if train_or_test == 'train':
self.exploration = max(0.1, self.exploration - 1e-6)
return act
def train(self, state, action, reward, next_state, isOver):
if self.global_step % self.update_target_steps == 0:
self.sync_target_network()
self.global_step += 1
action = np.expand_dims(action, -1)
self.exe.run(self.train_program,
feed={
'state': state.astype('float32'),
'action': action.astype('int32'),
'reward': reward,
'next_s': next_state.astype('float32'),
'isOver': isOver
})
def sync_target_network(self):
self.exe.run(self._sync_program)
fluid/DeepQNetwork/DoubleDQN_agent.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
fluid/DeepQNetwork/DuelingDQN_agent.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
fluid/DeepQNetwork/README.md 已删除 100644 → 0
浏览文件 @ 5ddb433e
[中文版](README_cn.md)
## Reproduce DQN, DoubleDQN, DuelingDQN model with Fluid version of PaddlePaddle
Based on PaddlePaddle's next-generation API Fluid, the DQN model of deep reinforcement learning is reproduced, and the same level of indicators of the paper is reproduced in the classic Atari game. The model receives the image of the game as input, and uses the end-to-end model to directly predict the next step. The repository contains the following three types of models:
+ DQN in
[Human-level Control Through Deep Reinforcement Learning](http://www.nature.com/nature/journal/v518/n7540/full/nature14236.html)
+ DoubleDQN in:
[Deep Reinforcement Learning with Double Q-Learning](https://www.aaai.org/ocs/index.php/AAAI/AAAI16/paper/viewPaper/12389)
+ DuelingDQN in:
[Dueling Network Architectures for Deep Reinforcement Learning](http://proceedings.mlr.press/v48/wangf16.html)
## Atari benchmark & performance
### Atari games introduction
Please see [here](https://gym.openai.com/envs/#atari) to know more about Atari game.
### Pong game result
The average game rewards that can be obtained for the three models as the number of training steps changes during the training are as follows(about 3 hours/1 Million steps):
<div align="center">
<img src="assets/dqn.png" width="600" height="300" alt="DQN result"></img>
</div>
## How to use
### Dependencies:
+ python2.7
+ gym
+ tqdm
+ opencv-python
+ paddlepaddle-gpu>=1.0.0
+ ale_python_interface
### Install Dependencies:
+ Install PaddlePaddle:
recommended to compile and install PaddlePaddle from source code
+ Install other dependencies:
```
pip install -r requirement.txt
pip install gym[atari]
```
Install ale_python_interface, please see [here](https://github.com/mgbellemare/Arcade-Learning-Environment).
### Start Training:
```
# To train a model for Pong game with gpu (use DQN model as default)
python train.py --rom ./rom_files/pong.bin --use_cuda
# To train a model for Pong with DoubleDQN
python train.py --rom ./rom_files/pong.bin --use_cuda --alg DoubleDQN
# To train a model for Pong with DuelingDQN
python train.py --rom ./rom_files/pong.bin --use_cuda --alg DuelingDQN
```
To train more games, you can install more rom files from [here](https://github.com/openai/atari-py/tree/master/atari_py/atari_roms).
### Start Testing:
```
# Play the game with saved best model and calculate the average rewards
python play.py --rom ./rom_files/pong.bin --use_cuda --model_path ./saved_model/DQN-pong
# Play the game with visualization
python play.py --rom ./rom_files/pong.bin --use_cuda --model_path ./saved_model/DQN-pong --viz 0.01
```
[Here](https://pan.baidu.com/s/1gIsbNw5V7tMeb74ojx-TMA) is saved models for Pong and Breakout games. You can use it to play the game directly.
fluid/DeepQNetwork/README_cn.md 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
fluid/DeepQNetwork/atari.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
fluid/DeepQNetwork/atari_wrapper.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
fluid/DeepQNetwork/expreplay.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
fluid/DeepQNetwork/play.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
fluid/DeepQNetwork/requirement.txt 已删除 100644 → 0
浏览文件 @ 5ddb433e
numpy
gym
tqdm
opencv-python
paddlepaddle-gpu==0.12.0
fluid/DeepQNetwork/train.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
fluid/PaddleCV/human_pose_estimation/demo.gif 已删除 100644 → 0
浏览文件 @ 5ddb433e
因为 它太大了无法显示 image diff 。你可以改为 查看blob。
fluid/PaddleCV/human_pose_estimation/test.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
fluid/PaddleCV/human_pose_estimation/train.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
fluid/PaddleCV/human_pose_estimation/val.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
fluid/PaddleCV/video_classification/README.md 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
fluid/PaddleCV/video_classification/eval.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
fluid/PaddleCV/video_classification/infer.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
fluid/PaddleCV/video_classification/reader.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
fluid/PaddleCV/video_classification/resnet.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
fluid/PaddleCV/video_classification/train.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
fluid/adversarial/README.md 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
fluid/adversarial/advbox/__init__.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
"""
A set of tools for generating adversarial example on paddle platform
"""
fluid/adversarial/advbox/adversary.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
fluid/adversarial/advbox/attacks/__init__.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
fluid/adversarial/advbox/attacks/base.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
fluid/adversarial/advbox/attacks/deepfool.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
fluid/adversarial/advbox/attacks/lbfgs.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
fluid/adversarial/advbox/attacks/saliency.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
fluid/adversarial/advbox/models/__init__.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
fluid/adversarial/advbox/models/base.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
fluid/adversarial/advbox/models/paddle.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
fluid/adversarial/tutorials/__init__.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
fluid/adversarial/tutorials/mnist_model.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
fluid/mnist/.run_ce.sh 已删除 100755 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
fluid/mnist/_ce.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
fluid/mnist/model.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
fluid/policy_gradient/README.md 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
fluid/policy_gradient/brain.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
fluid/policy_gradient/env.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
fluid/policy_gradient/images/PG_1.svg 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
fluid/policy_gradient/images/PG_2.svg 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
fluid/policy_gradient/images/PG_3.svg 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
fluid/policy_gradient/run.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/README.cn.md 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/README.md 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/conv_seq2seq/README.md 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/conv_seq2seq/beamsearch.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/conv_seq2seq/download.sh 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/conv_seq2seq/infer.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/conv_seq2seq/model.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/conv_seq2seq/preprocess.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/conv_seq2seq/reader.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/conv_seq2seq/train.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/ctr/README.cn.md 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/ctr/README.md 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/ctr/avazu_data_processer.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/ctr/dataset.md 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/ctr/images/lr_vs_dnn.jpg 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/ctr/images/wide_deep.png 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/ctr/infer.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/ctr/network_conf.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/ctr/reader.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/ctr/train.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/ctr/utils.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/deep_fm/README.cn.md 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/deep_fm/README.md 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/deep_fm/data/download.sh 已删除 100755 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/deep_fm/infer.py 已删除 100755 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/deep_fm/network_conf.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/deep_fm/preprocess.py 已删除 100755 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/deep_fm/reader.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/deep_fm/train.py 已删除 100755 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/dssm/README.cn.md 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/dssm/README.md 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/dssm/data/classification/test.txt 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/dssm/data/classification/train.txt 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/dssm/data/rank/test.txt 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/dssm/data/rank/train.txt 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/dssm/data/vocab.txt 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/dssm/images/dssm.jpg 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/dssm/images/dssm.png 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/dssm/images/dssm2.jpg 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/dssm/images/dssm2.png 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/dssm/images/dssm3.jpg 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/dssm/infer.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/dssm/network_conf.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/dssm/reader.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/dssm/train.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/dssm/utils.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/generate_chinese_poetry/README.md 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/generate_chinese_poetry/README_en.md 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/generate_chinese_poetry/generate.py 已删除 100755 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/generate_chinese_poetry/preprocess.py 已删除 100755 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/generate_chinese_poetry/reader.py 已删除 100755 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/generate_chinese_poetry/train.py 已删除 100755 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/generate_chinese_poetry/utils.py 已删除 100755 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/generate_sequence_by_rnn_lm/.gitignore 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/generate_sequence_by_rnn_lm/README.md 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/generate_sequence_by_rnn_lm/config.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/generate_sequence_by_rnn_lm/reader.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/generate_sequence_by_rnn_lm/train.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/generate_sequence_by_rnn_lm/utils.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/globally_normalized_reader/.gitignore 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/globally_normalized_reader/README.md 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/globally_normalized_reader/config.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/globally_normalized_reader/evaluate.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/globally_normalized_reader/infer.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/globally_normalized_reader/model.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/globally_normalized_reader/reader.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/globally_normalized_reader/train.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/globally_normalized_reader/vocab.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/hsigmoid/.gitignore 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/hsigmoid/README.md 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/hsigmoid/images/binary_tree.png 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/hsigmoid/images/network_conf.png 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/hsigmoid/images/path_to_1.png 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/hsigmoid/infer.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/hsigmoid/network_conf.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/hsigmoid/train.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/image_classification/README.md 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/image_classification/alexnet.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/image_classification/googlenet.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/image_classification/inception_v4.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/image_classification/infer.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/image_classification/reader.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/image_classification/resnet.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/image_classification/train.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/image_classification/vgg.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/image_classification/xception.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/ltr/README.md 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/ltr/README_en.md 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/ltr/images/LambdaRank_EN.png 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/ltr/images/lambdarank.jpg 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/ltr/images/learning_to_rank.jpg 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/ltr/images/ranknet.jpg 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/ltr/images/ranknet_en.png 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/ltr/images/search_engine_example.png 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/ltr/infer.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/ltr/lambda_rank.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/ltr/ranknet.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/ltr/train.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/mt_with_external_memory/README.md 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/mt_with_external_memory/data_utils.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/mt_with_external_memory/infer.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/mt_with_external_memory/model.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/mt_with_external_memory/train.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/nce_cost/.gitignore 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/nce_cost/README.md 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/nce_cost/images/network_conf.png 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/nce_cost/infer.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/nce_cost/network_conf.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/nce_cost/train.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/nested_sequence/README.md 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/nested_sequence/README_en.md 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/neural_qa/.gitignore 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/neural_qa/README.md 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/neural_qa/config.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/neural_qa/infer.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/neural_qa/network.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/neural_qa/reader.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/neural_qa/test/test_reader.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/neural_qa/test/trn_data.gz 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/neural_qa/train.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/neural_qa/utils.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/neural_qa/val_and_test.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/nmt_without_attention/README.cn.md 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/nmt_without_attention/README.md 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/nmt_without_attention/generate.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/nmt_without_attention/images/gru.png 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/nmt_without_attention/network_conf.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/nmt_without_attention/train.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/scene_text_recognition/README.md 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/scene_text_recognition/config.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/scene_text_recognition/decoder.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/scene_text_recognition/images/503.jpg 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/scene_text_recognition/images/504.jpg 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/scene_text_recognition/images/505.jpg 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/scene_text_recognition/images/ctc.png 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/scene_text_recognition/infer.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/scene_text_recognition/network_conf.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/scene_text_recognition/reader.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/scene_text_recognition/train.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/scene_text_recognition/utils.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/scheduled_sampling/README.md 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/scheduled_sampling/README_en.md 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/scheduled_sampling/generate.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/scheduled_sampling/images/decay.jpg 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/scheduled_sampling/network_conf.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/scheduled_sampling/reader.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/scheduled_sampling/train.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/scheduled_sampling/utils.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/sequence_tagging_for_ner/.gitignore 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/sequence_tagging_for_ner/README.md 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/sequence_tagging_for_ner/data/test 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/sequence_tagging_for_ner/data/train 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/sequence_tagging_for_ner/infer.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/sequence_tagging_for_ner/reader.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/sequence_tagging_for_ner/train.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/sequence_tagging_for_ner/utils.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/ssd/README.cn.md 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/ssd/README.md 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/ssd/config/__init__.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/ssd/config/pascal_voc_conf.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/ssd/data/label_list 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/ssd/data/prepare_voc_data.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/ssd/data_provider.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/ssd/eval.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/ssd/image_util.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/ssd/images/SSD300x300_map.png 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/ssd/images/ssd_network.png 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/ssd/images/vis_1.jpg 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/ssd/images/vis_2.jpg 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/ssd/images/vis_3.jpg 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/ssd/images/vis_4.jpg 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/ssd/infer.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/ssd/train.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/ssd/vgg_ssd_net.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/ssd/visual.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/text_classification/.gitignore 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/text_classification/README.md 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/text_classification/images/cnn_net.png 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/text_classification/images/dnn_net.png 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/text_classification/infer.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/text_classification/network_conf.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/text_classification/reader.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/text_classification/run.sh 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/text_classification/train.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/text_classification/utils.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/youtube_recall/README.cn.md 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/youtube_recall/README.md 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/youtube_recall/data/data.tar 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/youtube_recall/data_processor.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/youtube_recall/images/model_network.png 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/youtube_recall/infer.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/youtube_recall/infer_user.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/youtube_recall/item_vector.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/youtube_recall/network_conf.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/youtube_recall/reader.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/youtube_recall/train.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/youtube_recall/user_vector.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
legacy/youtube_recall/utils.py 已删除 100644 → 0
浏览文件 @ 5ddb433e
此差异已折叠。
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或